├── chatglm3 ├── __init__.py ├── configuration_chatglm.py ├── tokenization_chatglm.py └── quantization.py ├── examples ├── workflow_ipa.png ├── workflow_legacy.png ├── workflow_controlnet.png ├── workflow_inpainting.png ├── workflow_ipa_faceid.png ├── workflow_ipa_legacy.png ├── workflow_same_seed_test.png └── workflow_official_controlnet.png ├── configs ├── tokenizer │ ├── vocab.txt │ ├── tokenizer.model │ └── tokenizer_config.json └── text_encoder_config.json ├── clip_vit_336 └── config.json ├── pyproject.toml ├── .github └── workflows │ └── publish.yml ├── hook_comfyui_kolors_v1.py ├── .gitignore ├── mz_kolors_legacy.py ├── README.md ├── ComfyUI_IPAdapter_plus ├── CrossAttentionPatch.py ├── image_proj_models.py └── utils.py ├── __init__.py ├── mz_kolors_core.py ├── hook_comfyui_kolors_v2.py ├── LICENSE └── mz_kolors_utils.py /chatglm3/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/workflow_ipa.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/examples/workflow_ipa.png -------------------------------------------------------------------------------- /configs/tokenizer/vocab.txt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/configs/tokenizer/vocab.txt -------------------------------------------------------------------------------- /examples/workflow_legacy.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/examples/workflow_legacy.png -------------------------------------------------------------------------------- /configs/tokenizer/tokenizer.model: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/configs/tokenizer/tokenizer.model -------------------------------------------------------------------------------- /examples/workflow_controlnet.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/examples/workflow_controlnet.png -------------------------------------------------------------------------------- /examples/workflow_inpainting.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/examples/workflow_inpainting.png -------------------------------------------------------------------------------- /examples/workflow_ipa_faceid.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/examples/workflow_ipa_faceid.png -------------------------------------------------------------------------------- /examples/workflow_ipa_legacy.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/examples/workflow_ipa_legacy.png -------------------------------------------------------------------------------- /examples/workflow_same_seed_test.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/examples/workflow_same_seed_test.png -------------------------------------------------------------------------------- /examples/workflow_official_controlnet.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/MinusZoneAI/ComfyUI-Kolors-MZ/HEAD/examples/workflow_official_controlnet.png -------------------------------------------------------------------------------- /configs/tokenizer/tokenizer_config.json: -------------------------------------------------------------------------------- 1 | { 2 | "name_or_path": "THUDM/chatglm3-6b-base", 3 | "remove_space": false, 4 | "do_lower_case": false, 5 | "tokenizer_class": "ChatGLMTokenizer", 6 | "auto_map": { 7 | "AutoTokenizer": [ 8 | "tokenization_chatglm.ChatGLMTokenizer", 9 | null 10 | ] 11 | } 12 | } 13 | -------------------------------------------------------------------------------- /clip_vit_336/config.json: -------------------------------------------------------------------------------- 1 | { 2 | "attention_dropout": 0.0, 3 | "dropout": 0.0, 4 | "hidden_act": "quick_gelu", 5 | "hidden_size": 1024, 6 | "image_size": 336, 7 | "initializer_factor": 1.0, 8 | "initializer_range": 0.02, 9 | "intermediate_size": 4096, 10 | "layer_norm_eps": 1e-05, 11 | "model_type": "clip_vision_model", 12 | "num_attention_heads": 16, 13 | "num_channels": 3, 14 | "num_hidden_layers": 24, 15 | "patch_size": 14, 16 | "projection_dim": 768, 17 | "torch_dtype": "float32" 18 | } 19 | 20 | -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [project] 2 | name = "comfyui-kolors-mz" 3 | description = "Implementation of Kolors on ComfyUI\nReference from [a/https://github.com/kijai/ComfyUI-KwaiKolorsWrapper](https://github.com/kijai/ComfyUI-KwaiKolorsWrapper)\nUsing ComfyUI Native Sampling" 4 | version = "2.0.0" 5 | license = { file = "GPL-3.0 license" } 6 | 7 | [project.urls] 8 | Repository = "https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ" 9 | # Used by Comfy Registry https://comfyregistry.org 10 | 11 | [tool.comfy] 12 | PublisherId = "wailovet" 13 | DisplayName = "ComfyUI-Kolors-MZ" 14 | Icon = "" 15 | -------------------------------------------------------------------------------- /.github/workflows/publish.yml: -------------------------------------------------------------------------------- 1 | name: Publish to Comfy registry 2 | on: 3 | workflow_dispatch: 4 | push: 5 | branches: 6 | - main 7 | - master 8 | paths: 9 | - "pyproject.toml" 10 | 11 | permissions: 12 | issues: write 13 | 14 | jobs: 15 | publish-node: 16 | name: Publish Custom Node to registry 17 | runs-on: ubuntu-latest 18 | if: ${{ github.repository_owner == 'MinusZoneAI' }} 19 | steps: 20 | - name: Check out code 21 | uses: actions/checkout@v4 22 | - name: Publish Custom Node 23 | uses: Comfy-Org/publish-node-action@v1 24 | with: 25 | ## Add your own personal access token to your Github Repository secrets and reference it here. 26 | personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }} 27 | -------------------------------------------------------------------------------- /configs/text_encoder_config.json: -------------------------------------------------------------------------------- 1 | { 2 | "_name_or_path": "THUDM/chatglm3-6b-base", 3 | "model_type": "chatglm", 4 | "architectures": [ 5 | "ChatGLMModel" 6 | ], 7 | "auto_map": { 8 | "AutoConfig": "configuration_chatglm.ChatGLMConfig", 9 | "AutoModel": "modeling_chatglm.ChatGLMForConditionalGeneration", 10 | "AutoModelForCausalLM": "modeling_chatglm.ChatGLMForConditionalGeneration", 11 | "AutoModelForSeq2SeqLM": "modeling_chatglm.ChatGLMForConditionalGeneration", 12 | "AutoModelForSequenceClassification": "modeling_chatglm.ChatGLMForSequenceClassification" 13 | }, 14 | "add_bias_linear": false, 15 | "add_qkv_bias": true, 16 | "apply_query_key_layer_scaling": true, 17 | "apply_residual_connection_post_layernorm": false, 18 | "attention_dropout": 0.0, 19 | "attention_softmax_in_fp32": true, 20 | "bias_dropout_fusion": true, 21 | "ffn_hidden_size": 13696, 22 | "fp32_residual_connection": false, 23 | "hidden_dropout": 0.0, 24 | "hidden_size": 4096, 25 | "kv_channels": 128, 26 | "layernorm_epsilon": 1e-05, 27 | "multi_query_attention": true, 28 | "multi_query_group_num": 2, 29 | "num_attention_heads": 32, 30 | "num_layers": 28, 31 | "original_rope": true, 32 | "padded_vocab_size": 65024, 33 | "post_layer_norm": true, 34 | "rmsnorm": true, 35 | "seq_length": 32768, 36 | "use_cache": true, 37 | "torch_dtype": "float16", 38 | "transformers_version": "4.30.2", 39 | "tie_word_embeddings": false, 40 | "eos_token_id": 2, 41 | "pad_token_id": 0 42 | } -------------------------------------------------------------------------------- /chatglm3/configuration_chatglm.py: -------------------------------------------------------------------------------- 1 | from transformers import PretrainedConfig 2 | 3 | 4 | class ChatGLMConfig(PretrainedConfig): 5 | model_type = "chatglm" 6 | def __init__( 7 | self, 8 | num_layers=28, 9 | padded_vocab_size=65024, 10 | hidden_size=4096, 11 | ffn_hidden_size=13696, 12 | kv_channels=128, 13 | num_attention_heads=32, 14 | seq_length=2048, 15 | hidden_dropout=0.0, 16 | classifier_dropout=None, 17 | attention_dropout=0.0, 18 | layernorm_epsilon=1e-5, 19 | rmsnorm=True, 20 | apply_residual_connection_post_layernorm=False, 21 | post_layer_norm=True, 22 | add_bias_linear=False, 23 | add_qkv_bias=False, 24 | bias_dropout_fusion=True, 25 | multi_query_attention=False, 26 | multi_query_group_num=1, 27 | apply_query_key_layer_scaling=True, 28 | attention_softmax_in_fp32=True, 29 | fp32_residual_connection=False, 30 | quantization_bit=0, 31 | pre_seq_len=None, 32 | prefix_projection=False, 33 | **kwargs 34 | ): 35 | self.num_layers = num_layers 36 | self.vocab_size = padded_vocab_size 37 | self.padded_vocab_size = padded_vocab_size 38 | self.hidden_size = hidden_size 39 | self.ffn_hidden_size = ffn_hidden_size 40 | self.kv_channels = kv_channels 41 | self.num_attention_heads = num_attention_heads 42 | self.seq_length = seq_length 43 | self.hidden_dropout = hidden_dropout 44 | self.classifier_dropout = classifier_dropout 45 | self.attention_dropout = attention_dropout 46 | self.layernorm_epsilon = layernorm_epsilon 47 | self.rmsnorm = rmsnorm 48 | self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm 49 | self.post_layer_norm = post_layer_norm 50 | self.add_bias_linear = add_bias_linear 51 | self.add_qkv_bias = add_qkv_bias 52 | self.bias_dropout_fusion = bias_dropout_fusion 53 | self.multi_query_attention = multi_query_attention 54 | self.multi_query_group_num = multi_query_group_num 55 | self.apply_query_key_layer_scaling = apply_query_key_layer_scaling 56 | self.attention_softmax_in_fp32 = attention_softmax_in_fp32 57 | self.fp32_residual_connection = fp32_residual_connection 58 | self.quantization_bit = quantization_bit 59 | self.pre_seq_len = pre_seq_len 60 | self.prefix_projection = prefix_projection 61 | super().__init__(**kwargs) 62 | -------------------------------------------------------------------------------- /hook_comfyui_kolors_v1.py: -------------------------------------------------------------------------------- 1 | from comfy.model_detection import * 2 | import comfy.model_detection as model_detection 3 | import comfy.supported_models 4 | 5 | 6 | class Kolors(comfy.supported_models.SDXL): 7 | unet_config = { 8 | "model_channels": 320, 9 | "use_linear_in_transformer": True, 10 | "transformer_depth": [0, 0, 2, 2, 10, 10], 11 | "context_dim": 2048, 12 | "adm_in_channels": 5632, 13 | "use_temporal_attention": False, 14 | } 15 | 16 | 17 | def kolors_unet_config_from_diffusers_unet(state_dict, dtype=None): 18 | match = {} 19 | transformer_depth = [] 20 | 21 | attn_res = 1 22 | down_blocks = count_blocks(state_dict, "down_blocks.{}") 23 | for i in range(down_blocks): 24 | attn_blocks = count_blocks( 25 | state_dict, "down_blocks.{}.attentions.".format(i) + '{}') 26 | res_blocks = count_blocks( 27 | state_dict, "down_blocks.{}.resnets.".format(i) + '{}') 28 | for ab in range(attn_blocks): 29 | transformer_count = count_blocks( 30 | state_dict, "down_blocks.{}.attentions.{}.transformer_blocks.".format(i, ab) + '{}') 31 | transformer_depth.append(transformer_count) 32 | if transformer_count > 0: 33 | match["context_dim"] = state_dict["down_blocks.{}.attentions.{}.transformer_blocks.0.attn2.to_k.weight".format( 34 | i, ab)].shape[1] 35 | 36 | attn_res *= 2 37 | if attn_blocks == 0: 38 | for i in range(res_blocks): 39 | transformer_depth.append(0) 40 | 41 | match["transformer_depth"] = transformer_depth 42 | 43 | match["model_channels"] = state_dict["conv_in.weight"].shape[0] 44 | match["in_channels"] = state_dict["conv_in.weight"].shape[1] 45 | match["adm_in_channels"] = None 46 | if "class_embedding.linear_1.weight" in state_dict: 47 | match["adm_in_channels"] = state_dict["class_embedding.linear_1.weight"].shape[1] 48 | elif "add_embedding.linear_1.weight" in state_dict: 49 | match["adm_in_channels"] = state_dict["add_embedding.linear_1.weight"].shape[1] 50 | 51 | Kolors = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 52 | 'num_classes': 'sequential', 'adm_in_channels': 5632, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 53 | 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10, 54 | 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], 55 | 'use_temporal_attention': False, 'use_temporal_resblock': False} 56 | 57 | supported_models = [Kolors] 58 | 59 | for unet_config in supported_models: 60 | matches = True 61 | for k in match: 62 | if match[k] != unet_config[k]: 63 | print("key {} does not match".format( 64 | k), match[k], "||", unet_config[k]) 65 | matches = False 66 | break 67 | if matches: 68 | return convert_config(unet_config) 69 | return None 70 | 71 | 72 | class apply_kolors: 73 | def __enter__(self): 74 | import comfy.supported_models 75 | self.old_supported_models = comfy.supported_models.models 76 | comfy.supported_models.models = [Kolors] 77 | 78 | self.old_unet_config_from_diffusers_unet = model_detection.unet_config_from_diffusers_unet 79 | model_detection.unet_config_from_diffusers_unet = kolors_unet_config_from_diffusers_unet 80 | 81 | def __exit__(self, type, value, traceback): 82 | model_detection.unet_config_from_diffusers_unet = self.old_unet_config_from_diffusers_unet 83 | 84 | import comfy.supported_models 85 | comfy.supported_models.models = self.old_supported_models 86 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Byte-compiled / optimized / DLL files 2 | __pycache__/ 3 | *.py[cod] 4 | *$py.class 5 | 6 | # C extensions 7 | *.so 8 | 9 | # Distribution / packaging 10 | .Python 11 | build/ 12 | develop-eggs/ 13 | dist/ 14 | downloads/ 15 | eggs/ 16 | .eggs/ 17 | lib/ 18 | lib64/ 19 | parts/ 20 | sdist/ 21 | var/ 22 | wheels/ 23 | share/python-wheels/ 24 | *.egg-info/ 25 | .installed.cfg 26 | *.egg 27 | MANIFEST 28 | 29 | # PyInstaller 30 | # Usually these files are written by a python script from a template 31 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 32 | *.manifest 33 | *.spec 34 | 35 | # Installer logs 36 | pip-log.txt 37 | pip-delete-this-directory.txt 38 | 39 | # Unit test / coverage reports 40 | htmlcov/ 41 | .tox/ 42 | .nox/ 43 | .coverage 44 | .coverage.* 45 | .cache 46 | nosetests.xml 47 | coverage.xml 48 | *.cover 49 | *.py,cover 50 | .hypothesis/ 51 | .pytest_cache/ 52 | cover/ 53 | 54 | # Translations 55 | *.mo 56 | *.pot 57 | 58 | # Django stuff: 59 | *.log 60 | local_settings.py 61 | db.sqlite3 62 | db.sqlite3-journal 63 | 64 | # Flask stuff: 65 | instance/ 66 | .webassets-cache 67 | 68 | # Scrapy stuff: 69 | .scrapy 70 | 71 | # Sphinx documentation 72 | docs/_build/ 73 | 74 | # PyBuilder 75 | .pybuilder/ 76 | target/ 77 | 78 | # Jupyter Notebook 79 | .ipynb_checkpoints 80 | 81 | # IPython 82 | profile_default/ 83 | ipython_config.py 84 | 85 | # pyenv 86 | # For a library or package, you might want to ignore these files since the code is 87 | # intended to run in multiple environments; otherwise, check them in: 88 | # .python-version 89 | 90 | # pipenv 91 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 92 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 93 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 94 | # install all needed dependencies. 95 | #Pipfile.lock 96 | 97 | # poetry 98 | # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. 99 | # This is especially recommended for binary packages to ensure reproducibility, and is more 100 | # commonly ignored for libraries. 101 | # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control 102 | #poetry.lock 103 | 104 | # pdm 105 | # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. 106 | #pdm.lock 107 | # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it 108 | # in version control. 109 | # https://pdm.fming.dev/latest/usage/project/#working-with-version-control 110 | .pdm.toml 111 | .pdm-python 112 | .pdm-build/ 113 | 114 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm 115 | __pypackages__/ 116 | 117 | # Celery stuff 118 | celerybeat-schedule 119 | celerybeat.pid 120 | 121 | # SageMath parsed files 122 | *.sage.py 123 | 124 | # Environments 125 | .env 126 | .venv 127 | env/ 128 | venv/ 129 | ENV/ 130 | env.bak/ 131 | venv.bak/ 132 | 133 | # Spyder project settings 134 | .spyderproject 135 | .spyproject 136 | 137 | # Rope project settings 138 | .ropeproject 139 | 140 | # mkdocs documentation 141 | /site 142 | 143 | # mypy 144 | .mypy_cache/ 145 | .dmypy.json 146 | dmypy.json 147 | 148 | # Pyre type checker 149 | .pyre/ 150 | 151 | # pytype static type analyzer 152 | .pytype/ 153 | 154 | # Cython debug symbols 155 | cython_debug/ 156 | 157 | # PyCharm 158 | # JetBrains specific template is maintained in a separate JetBrains.gitignore that can 159 | # be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore 160 | # and can be added to the global gitignore or merged into this file. For a more nuclear 161 | # option (not recommended) you can uncomment the following to ignore the entire idea folder. 162 | #.idea/ 163 | exclude.txt 164 | copytoww.bat 165 | -------------------------------------------------------------------------------- /mz_kolors_legacy.py: -------------------------------------------------------------------------------- 1 | 2 | 3 | import gc 4 | import json 5 | import os 6 | import random 7 | import re 8 | 9 | import torch 10 | import folder_paths 11 | import comfy.model_management as mm 12 | from . import mz_kolors_core 13 | 14 | 15 | def MZ_ChatGLM3TextEncode_call(args): 16 | 17 | text = args.get("text") 18 | chatglm3_model = args.get("chatglm3_model") 19 | 20 | prompt_embeds, pooled_output = mz_kolors_core.chatglm3_text_encode( 21 | chatglm3_model, 22 | text, 23 | ) 24 | 25 | from torch import nn 26 | hid_proj: nn.Linear = args.get("hid_proj") 27 | 28 | if hid_proj.weight.dtype != prompt_embeds.dtype: 29 | with torch.cuda.amp.autocast(dtype=hid_proj.weight.dtype): 30 | prompt_embeds = hid_proj(prompt_embeds) 31 | else: 32 | prompt_embeds = hid_proj(prompt_embeds) 33 | 34 | return ([[ 35 | prompt_embeds, 36 | {"pooled_output": pooled_output}, 37 | ]], ) 38 | 39 | 40 | def load_unet_state_dict(sd): # load unet in diffusers or regular format 41 | from comfy import model_management, model_detection 42 | import comfy.utils 43 | 44 | # Allow loading unets from checkpoint files 45 | checkpoint = False 46 | diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd) 47 | temp_sd = comfy.utils.state_dict_prefix_replace( 48 | sd, {diffusion_model_prefix: ""}, filter_keys=True) 49 | if len(temp_sd) > 0: 50 | sd = temp_sd 51 | checkpoint = True 52 | 53 | parameters = comfy.utils.calculate_parameters(sd) 54 | unet_dtype = model_management.unet_dtype(model_params=parameters) 55 | load_device = model_management.get_torch_device() 56 | 57 | from torch import nn 58 | hid_proj: nn.Linear = None 59 | if True: 60 | model_config = model_detection.model_config_from_diffusers_unet(sd) 61 | if model_config is None: 62 | return None 63 | 64 | diffusers_keys = comfy.utils.unet_to_diffusers( 65 | model_config.unet_config) 66 | 67 | new_sd = {} 68 | for k in diffusers_keys: 69 | if k in sd: 70 | new_sd[diffusers_keys[k]] = sd.pop(k) 71 | else: 72 | print("{} {}".format(diffusers_keys[k], k)) 73 | 74 | encoder_hid_proj_weight = sd.pop("encoder_hid_proj.weight") 75 | encoder_hid_proj_bias = sd.pop("encoder_hid_proj.bias") 76 | hid_proj = nn.Linear( 77 | encoder_hid_proj_weight.shape[1], encoder_hid_proj_weight.shape[0]) 78 | hid_proj.weight.data = encoder_hid_proj_weight 79 | hid_proj.bias.data = encoder_hid_proj_bias 80 | hid_proj = hid_proj.to(load_device) 81 | 82 | offload_device = model_management.unet_offload_device() 83 | unet_dtype = model_management.unet_dtype( 84 | model_params=parameters, supported_dtypes=model_config.supported_inference_dtypes) 85 | manual_cast_dtype = model_management.unet_manual_cast( 86 | unet_dtype, load_device, model_config.supported_inference_dtypes) 87 | model_config.set_inference_dtype(unet_dtype, manual_cast_dtype) 88 | model = model_config.get_model(new_sd, "") 89 | model = model.to(offload_device) 90 | model.load_model_weights(new_sd, "") 91 | left_over = sd.keys() 92 | if len(left_over) > 0: 93 | print("left over keys in unet: {}".format(left_over)) 94 | return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device), hid_proj 95 | 96 | 97 | def MZ_KolorsUNETLoader_call(kwargs): 98 | 99 | from . import hook_comfyui_kolors_v1 100 | with hook_comfyui_kolors_v1.apply_kolors(): 101 | unet_name = kwargs.get("unet_name") 102 | unet_path = folder_paths.get_full_path("unet", unet_name) 103 | import comfy.utils 104 | sd = comfy.utils.load_torch_file(unet_path) 105 | model, hid_proj = load_unet_state_dict(sd) 106 | if model is None: 107 | raise RuntimeError( 108 | "ERROR: Could not detect model type of: {}".format(unet_path)) 109 | return (model, hid_proj) 110 | 111 | 112 | def MZ_FakeCond_call(kwargs): 113 | import torch 114 | cond = torch.zeros(2, 256, 4096) 115 | pool = torch.zeros(2, 4096) 116 | 117 | dtype = kwargs.get("dtype") 118 | if dtype == "fp16": 119 | print("fp16") 120 | cond = cond.half() 121 | pool = pool.half() 122 | elif dtype == "bf16": 123 | print("bf16") 124 | cond = cond.bfloat16() 125 | pool = pool.bfloat16() 126 | else: 127 | print("fp32") 128 | cond = cond.float() 129 | pool = pool.float() 130 | 131 | return ([[ 132 | cond, 133 | {"pooled_output": pool}, 134 | ]],) 135 | 136 | 137 | NODE_CLASS_MAPPINGS = { 138 | } 139 | 140 | 141 | NODE_DISPLAY_NAME_MAPPINGS = { 142 | } 143 | 144 | AUTHOR_NAME = "MinusZone" 145 | CATEGORY_NAME = f"{AUTHOR_NAME} - Kolors" 146 | 147 | 148 | class MZ_ChatGLM3TextEncode: 149 | @classmethod 150 | def INPUT_TYPES(s): 151 | return { 152 | "required": { 153 | "chatglm3_model": ("CHATGLM3MODEL", ), 154 | "text": ("STRING", {"multiline": True, "dynamicPrompts": True}), 155 | "hid_proj": ("TorchLinear", ), 156 | } 157 | } 158 | 159 | RETURN_TYPES = ("CONDITIONING",) 160 | 161 | FUNCTION = "encode" 162 | CATEGORY = CATEGORY_NAME + "/Legacy" 163 | 164 | def encode(self, **kwargs): 165 | return MZ_ChatGLM3TextEncode_call(kwargs) 166 | 167 | 168 | NODE_CLASS_MAPPINGS["MZ_ChatGLM3"] = MZ_ChatGLM3TextEncode 169 | NODE_DISPLAY_NAME_MAPPINGS[ 170 | "MZ_ChatGLM3"] = f"{AUTHOR_NAME} - ChatGLM3TextEncode" 171 | 172 | 173 | class MZ_KolorsUNETLoader(): 174 | @classmethod 175 | def INPUT_TYPES(s): 176 | return {"required": { 177 | "unet_name": (folder_paths.get_filename_list("unet"), ), 178 | }} 179 | 180 | RETURN_TYPES = ("MODEL", "TorchLinear") 181 | 182 | RETURN_NAMES = ("model", "hid_proj") 183 | 184 | FUNCTION = "load_unet" 185 | 186 | CATEGORY = CATEGORY_NAME + "/Legacy" 187 | 188 | def load_unet(self, **kwargs): 189 | return MZ_KolorsUNETLoader_call(kwargs) 190 | 191 | 192 | NODE_CLASS_MAPPINGS["MZ_KolorsUNETLoader"] = MZ_KolorsUNETLoader 193 | NODE_DISPLAY_NAME_MAPPINGS[ 194 | "MZ_KolorsUNETLoader"] = f"{AUTHOR_NAME} - Kolors UNET Loader" 195 | 196 | 197 | class MZ_FakeCond: 198 | @classmethod 199 | def INPUT_TYPES(s): 200 | return { 201 | "required": { 202 | "seed": ("INT", {"default": 0}), 203 | "dtype": ([ 204 | "fp32", 205 | "fp16", 206 | "bf16", 207 | ],), 208 | } 209 | } 210 | 211 | RETURN_TYPES = ("CONDITIONING", ) 212 | RETURN_NAMES = ("prompt", ) 213 | FUNCTION = "encode" 214 | CATEGORY = CATEGORY_NAME 215 | 216 | def encode(self, **kwargs): 217 | return MZ_FakeCond_call(kwargs) 218 | 219 | 220 | try: 221 | if os.environ.get("MZ_DEV", None) is not None: 222 | NODE_CLASS_MAPPINGS["MZ_FakeCond"] = MZ_FakeCond 223 | NODE_DISPLAY_NAME_MAPPINGS[ 224 | "MZ_FakeCond"] = f"{AUTHOR_NAME} - FakeCond" 225 | except ImportError: 226 | pass 227 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ![image](./examples/workflow_ipa.png) 2 | 3 | ## Recent changes 4 | * [2024-08-02] 支持faceid,新增一些相关节点,工作流见examples/workflow_ipa_faceid.png 5 | * [2024-07-27] 新增MZ_KolorsControlNetLoader节点,用于加载可图ControlNet官方模型 6 | * [2024-07-26] 新增MZ_ApplySDXLSamplingSettings节点,用于V2版本重新回到SDXL的scheduler配置. 7 | ![image](https://github.com/user-attachments/assets/8c3be6bf-4744-478f-8660-4842a4558a1f) 8 | 9 | * [2024-07-25] 修正sampling_settings,参数来自 [scheduler_config.json](https://huggingface.co/Kwai-Kolors/Kolors/blob/main/scheduler/scheduler_config.json),仅V2生效 10 | * [2024-07-21] 感谢来自yiwangsimple对Mac修复和测试的分支 https://github.com/yiwangsimple/ComfyUI-Kolors-MZ 11 | * [2024-07-21] 新增MZ_ChatGLM3TextEncodeAdvanceV2节点 12 | * [2024-07-18] IPA相关节点已在ComfyUI_IPAdapter_plus中支持 13 | * [2024-07-17] 新增支持IPAdapter_plus的加载器和高级应用节点 MZ_KolorsCLIPVisionLoader,MZ_IPAdapterModelLoaderKolors,MZ_IPAdapterAdvancedKolors 14 | * [2024-07-14] 删除自动兼容ControlNet, 新增MZ_KolorsControlNetPatch节点 15 | ![image](https://github.com/user-attachments/assets/73ae6447-c69d-4781-9c66-94e0029709ed) 16 | 17 | ## ComfyUI上Kolors的实现 18 | 19 | 参考自 https://github.com/kijai/ComfyUI-KwaiKolorsWrapper 20 | 21 | 使用ComfyUI原生采样 22 | 23 | 工作流在examples/workflow.png中获取 24 | 25 | ### UNET模型下载 26 | unet模型放置在 models/unet/ 文件夹下 27 | 28 | 模型主页: https://huggingface.co/Kwai-Kolors/Kolors 29 | 30 | 下载地址: https://huggingface.co/Kwai-Kolors/Kolors/resolve/main/unet/diffusion_pytorch_model.fp16.safetensors 31 | 32 | 33 | ### ChatGLM3模型下载 34 | chatglm3放置在 models/LLM/ 文件夹下 35 | 36 | 模型主页: https://huggingface.co/Kijai/ChatGLM3-safetensors 37 | 38 | 下载地址: https://huggingface.co/Kijai/ChatGLM3-safetensors/resolve/main/chatglm3-fp16.safetensors 39 | 40 | 41 | ## IPAdapter实现推荐使用 [cubiq/ComfyUI_IPAdapter_plus](https://github.com/cubiq/ComfyUI_IPAdapter_plus) 42 | faceid相关节点已经在其中支持,IPAdapter实现需要更新ComfyUI_IPAdapter_plus到最新版本 43 | ### IPAdapter工作流 44 | https://github.com/cubiq/ComfyUI_IPAdapter_plus/blob/main/examples/ipadapter_kolors.json 45 | 46 | ### IPAdapter_FaceIDv2工作流 47 | https://github.com/cubiq/ComfyUI_IPAdapter_plus/blob/main/examples/IPAdapter_FaceIDv2_Kolors.json 48 | 49 | 50 | ### 官方IP-Adapter-Plus模型下载地址 51 | 模型主页: https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-Plus 52 | 53 | https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-Plus/resolve/main/ip_adapter_plus_general.bin 下载至 models/ipadapter/ 54 | 55 | https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-Plus/resolve/main/image_encoder/pytorch_model.bin 下载至 models/clip_vision/ 56 | 57 | ### 官方Kolors-IP-Adapter-FaceID-Plus模型下载地址 58 | 模型主页: https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-FaceID-Plus 59 | 60 | https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-FaceID-Plus/resolve/main/ipa-faceid-plus.bin 下载至 models/ipadapter/ 61 | 62 | https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-FaceID-Plus/resolve/main/clip-vit-large-patch14-336/pytorch_model.bin 下载至 models/clip_vision/ 63 | 64 | https://huggingface.co/MonsterMMORPG/tools/resolve/main/antelopev2.zip 下载并解压至 models/insightface/models/antelopev2/*.onnx 65 | 66 | ### 官方ControlNet模型下载地址 67 | 模型主页(Depth): https://huggingface.co/Kwai-Kolors/Kolors-ControlNet-Depth 68 | 69 | 模型主页(Canny): https://huggingface.co/Kwai-Kolors/Kolors-ControlNet-Canny 70 | 71 | 模型主页(Pose): https://huggingface.co/Kwai-Kolors/Kolors-ControlNet-Pose 72 | 73 | ### Kolors-Inpainting模型下载地址 74 | 模型主页: https://huggingface.co/Kwai-Kolors/Kolors-Inpainting 75 | 76 | https://huggingface.co/Kwai-Kolors/Kolors-Inpainting/resolve/main/unet/diffusion_pytorch_model.safetensors 下载至 models/unet/ 77 | 78 | 79 | 80 | ## Implementation of Kolors on ComfyUI 81 | 82 | Reference from https://github.com/kijai/ComfyUI-KwaiKolorsWrapper 83 | 84 | Using ComfyUI Native Sampling 85 | 86 | The workflow is obtained in examples/workflow.png 87 | 88 | 89 | ### UNET model download 90 | The unet model is placed in the models/unet/ folder 91 | 92 | Model homepage: https://huggingface.co/Kwai-Kolors/Kolors 93 | 94 | Download link: 95 | https://huggingface.co/Kwai-Kolors/Kolors/resolve/main/unet/diffusion_pytorch_model.fp16.safetensors 96 | 97 | 98 | ### ChatGLM3 model download 99 | The chatglm3 is placed in the models/LLM/ folder 100 | 101 | Model homepage: https://huggingface.co/Kijai/ChatGLM3-safetensors 102 | 103 | Download link: 104 | https://huggingface.co/Kijai/ChatGLM3-safetensors/resolve/main/chatglm3-fp16.safetensors 105 | 106 | 107 | ## IPAdapter implementation is recommended to use [cubiq/ComfyUI_IPAdapter_plus](https://github.com/cubiq/ComfyUI_IPAdapter_plus) 108 | The faceid related nodes have been supported in it, and the IPAdapter implementation needs to update ComfyUI_IPAdapter_plus to the latest version 109 | ### IPAdapter workflow 110 | https://github.com/cubiq/ComfyUI_IPAdapter_plus/blob/main/examples/ipadapter_kolors.json 111 | 112 | ### IPAdapter_FaceIDv2 workflow 113 | https://github.com/cubiq/ComfyUI_IPAdapter_plus/blob/main/examples/IPAdapter_FaceIDv2_Kolors.json 114 | 115 | 116 | ### Official IP-Adapter-Plus model download link 117 | 118 | Model homepage: https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-Plus 119 | 120 | https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-Plus/resolve/main/ip_adapter_plus_general.bin Download to models/ipadapter/ 121 | 122 | https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-Plus/resolve/main/image_encoder/pytorch_model.bin Download to models/clip_vision/ 123 | 124 | 125 | ### Official Kolors-IP-Adapter-FaceID-Plus model download link 126 | Model homepage: https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-FaceID-Plus 127 | 128 | https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-FaceID-Plus/resolve/main/ipa-faceid-plus.bin Download to models/ipadapter/ 129 | 130 | https://huggingface.co/Kwai-Kolors/Kolors-IP-Adapter-FaceID-Plus/resolve/main/clip-vit-large-patch14-336/pytorch_model.bin Download to models/clip_vision/ 131 | 132 | https://huggingface.co/MonsterMMORPG/tools/resolve/main/antelopev2.zip Download and unzip to models/insightface/models/antelopev2/*.onnx 133 | 134 | ### Official ControlNet model download link 135 | Model homepage(Depth): https://huggingface.co/Kwai-Kolors/Kolors-ControlNet-Depth 136 | 137 | Model homepage(Canny): https://huggingface.co/Kwai-Kolors/Kolors-ControlNet-Canny 138 | 139 | Model homepage(Pose): https://huggingface.co/Kwai-Kolors/Kolors-ControlNet-Pose 140 | 141 | ### Kolors-Inpainting model download link 142 | Model homepage: https://huggingface.co/Kwai-Kolors/Kolors-Inpainting 143 | 144 | https://huggingface.co/Kwai-Kolors/Kolors-Inpainting/resolve/main/unet/diffusion_pytorch_model.safetensors Download to models/unet/ 145 | 146 | 147 | ## 使用ComfyUI-KwaiKolorsWrapper在相同种子下测试结果 (Testing results with the same seed using ComfyUI-KwaiKolorsWrapper) 148 | 测试工作流来自examples/workflow_same_seed_test.png 149 | 150 | The test workflow comes from examples/workflow_same_seed_test.png 151 | 152 | ![image](./examples/workflow_same_seed_test.png) 153 | 154 | ## FAQ 155 | 加载模型时出现的错误 156 | + 目前kolors有两个版本, 一种是unet类型采用unet加载器, 一种是放checkpoints文件夹用KolorsCheckpointLoaderSimple加载器, 如果你的模型来自huggingface的unet文件夹, 优先尝试使用unet加载器 (Currently, there are two versions of kolors, one is unet type using unet loader, and the other is placed in the checkpoints folder using KolorsCheckpointLoaderSimple loader. If your model comes from the huggingface unet folder, try to use the unet loader first) 157 | + 如果你无法确定模型类型, 那就都试一下 (If you are not sure about the model type, try both) 158 | 159 | Mac用户无法使用(Mac users cannot use) 160 | + Mac用户可移步至[ComfyUI-Kolors-MZ](https://github.com/yiwangsimple/ComfyUI-Kolors-MZ) (Mac users can go to [ComfyUI-Kolors-MZ](https://github.com/yiwangsimple/ComfyUI-Kolors-MZ) ) 161 | 162 | 和IPAdapter有关的错误(Errors related to IPAdapter) 163 | + 确保ComfyUI本体和ComfyUI_IPAdapter_plus已经更新到最新版本(Make sure ComfyUI ontology and ComfyUI_IPAdapter_plus are updated to the latest version) 164 | 165 | name 'round_up' is not defined 166 | + 参考:https://github.com/THUDM/ChatGLM2-6B/issues/272#issuecomment-1632164243 , 使用 pip install cpm_kernels 或者 pip install -U cpm_kernels 更新 cpm_kernels 167 | 168 | module 'comfy.model_detection' has no attribute 'unet_prefix_from_state_dict' 169 | + 更新ComfyUI本体到最新版本(Update ComfyUI ontology to the latest version) 170 | 171 | RuntimeError: Only Tensors of floating point dtype can require gradients 172 | + 尝试使用fp16版本的模型: https://huggingface.co/Kijai/ChatGLM3-safetensors/blob/main/chatglm3-fp16.safetensors 173 | 174 | Error occurred when executing MZ_ChatGLM3Loader: 'ChatGLMModel' object has no attribute 'transformer' 175 | + 检查ChatGLM3Loader节点选择的模型是否已经正确下载 176 | 177 | 178 | ## Credits 179 | 180 | - [Kolors](https://github.com/Kwai-Kolors/Kolors) 181 | - [ComfyUI](https://github.com/comfyanonymous/ComfyUI) 182 | - [ComfyUI_IPAdapter_plus](https://github.com/cubiq/ComfyUI_IPAdapter_plus) 183 | 184 | ## Star History 185 | 186 | 187 | 188 | 189 | 190 | Star History Chart 191 | 192 | 193 | 194 | 195 | ## Contact 196 | - 微信Wechat: minrszone 197 | - Bilibili: [minus_zone](https://space.bilibili.com/5950992) 198 | - 小红书: [MinusZoneAI](https://www.xiaohongshu.com/user/profile/5f072e990000000001005472) 199 | 200 | ## Stargazers 201 | [![Stargazers repo roster for @MinusZoneAI/ComfyUI-Kolors-MZ](https://reporoster.com/stars/MinusZoneAI/ComfyUI-Kolors-MZ)](https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ/stargazers) 202 | 203 | 204 | 205 | 206 | -------------------------------------------------------------------------------- /ComfyUI_IPAdapter_plus/CrossAttentionPatch.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import math 3 | import torch.nn.functional as F 4 | from comfy.ldm.modules.attention import optimized_attention 5 | from .utils import tensor_to_size 6 | 7 | class Attn2Replace: 8 | def __init__(self, callback=None, **kwargs): 9 | self.callback = [callback] 10 | self.kwargs = [kwargs] 11 | 12 | def add(self, callback, **kwargs): 13 | self.callback.append(callback) 14 | self.kwargs.append(kwargs) 15 | 16 | for key, value in kwargs.items(): 17 | setattr(self, key, value) 18 | 19 | def __call__(self, q, k, v, extra_options): 20 | dtype = q.dtype 21 | out = optimized_attention(q, k, v, extra_options["n_heads"]) 22 | sigma = extra_options["sigmas"].detach().cpu()[0].item() if 'sigmas' in extra_options else 999999999.9 23 | 24 | for i, callback in enumerate(self.callback): 25 | if sigma <= self.kwargs[i]["sigma_start"] and sigma >= self.kwargs[i]["sigma_end"]: 26 | out = out + callback(out, q, k, v, extra_options, **self.kwargs[i]) 27 | 28 | return out.to(dtype=dtype) 29 | 30 | def ipadapter_attention(out, q, k, v, extra_options, module_key='', ipadapter=None, weight=1.0, cond=None, cond_alt=None, uncond=None, weight_type="linear", mask=None, sigma_start=0.0, sigma_end=1.0, unfold_batch=False, embeds_scaling='V only', **kwargs): 31 | dtype = q.dtype 32 | cond_or_uncond = extra_options["cond_or_uncond"] 33 | block_type = extra_options["block"][0] 34 | #block_id = extra_options["block"][1] 35 | t_idx = extra_options["transformer_index"] 36 | layers = 11 if '101_to_k_ip' in ipadapter.ip_layers.to_kvs else 16 37 | k_key = module_key + "_to_k_ip" 38 | v_key = module_key + "_to_v_ip" 39 | 40 | # extra options for AnimateDiff 41 | ad_params = extra_options['ad_params'] if "ad_params" in extra_options else None 42 | 43 | b = q.shape[0] 44 | seq_len = q.shape[1] 45 | batch_prompt = b // len(cond_or_uncond) 46 | _, _, oh, ow = extra_options["original_shape"] 47 | 48 | if weight_type == 'ease in': 49 | weight = weight * (0.05 + 0.95 * (1 - t_idx / layers)) 50 | elif weight_type == 'ease out': 51 | weight = weight * (0.05 + 0.95 * (t_idx / layers)) 52 | elif weight_type == 'ease in-out': 53 | weight = weight * (0.05 + 0.95 * (1 - abs(t_idx - (layers/2)) / (layers/2))) 54 | elif weight_type == 'reverse in-out': 55 | weight = weight * (0.05 + 0.95 * (abs(t_idx - (layers/2)) / (layers/2))) 56 | elif weight_type == 'weak input' and block_type == 'input': 57 | weight = weight * 0.2 58 | elif weight_type == 'weak middle' and block_type == 'middle': 59 | weight = weight * 0.2 60 | elif weight_type == 'weak output' and block_type == 'output': 61 | weight = weight * 0.2 62 | elif weight_type == 'strong middle' and (block_type == 'input' or block_type == 'output'): 63 | weight = weight * 0.2 64 | elif isinstance(weight, dict): 65 | if t_idx not in weight: 66 | return 0 67 | 68 | if weight_type == "style transfer precise": 69 | if layers == 11 and t_idx == 3: 70 | uncond = cond 71 | cond = cond * 0 72 | elif layers == 16 and (t_idx == 4 or t_idx == 5): 73 | uncond = cond 74 | cond = cond * 0 75 | elif weight_type == "composition precise": 76 | if layers == 11 and t_idx != 3: 77 | uncond = cond 78 | cond = cond * 0 79 | elif layers == 16 and (t_idx != 4 and t_idx != 5): 80 | uncond = cond 81 | cond = cond * 0 82 | 83 | weight = weight[t_idx] 84 | 85 | if cond_alt is not None and t_idx in cond_alt: 86 | cond = cond_alt[t_idx] 87 | del cond_alt 88 | 89 | if unfold_batch: 90 | # Check AnimateDiff context window 91 | if ad_params is not None and ad_params["sub_idxs"] is not None: 92 | if isinstance(weight, torch.Tensor): 93 | weight = tensor_to_size(weight, ad_params["full_length"]) 94 | weight = torch.Tensor(weight[ad_params["sub_idxs"]]) 95 | if torch.all(weight == 0): 96 | return 0 97 | weight = weight.repeat(len(cond_or_uncond), 1, 1) # repeat for cond and uncond 98 | elif weight == 0: 99 | return 0 100 | 101 | # if image length matches or exceeds full_length get sub_idx images 102 | if cond.shape[0] >= ad_params["full_length"]: 103 | cond = torch.Tensor(cond[ad_params["sub_idxs"]]) 104 | uncond = torch.Tensor(uncond[ad_params["sub_idxs"]]) 105 | # otherwise get sub_idxs images 106 | else: 107 | cond = tensor_to_size(cond, ad_params["full_length"]) 108 | uncond = tensor_to_size(uncond, ad_params["full_length"]) 109 | cond = cond[ad_params["sub_idxs"]] 110 | uncond = uncond[ad_params["sub_idxs"]] 111 | else: 112 | if isinstance(weight, torch.Tensor): 113 | weight = tensor_to_size(weight, batch_prompt) 114 | if torch.all(weight == 0): 115 | return 0 116 | weight = weight.repeat(len(cond_or_uncond), 1, 1) # repeat for cond and uncond 117 | elif weight == 0: 118 | return 0 119 | 120 | cond = tensor_to_size(cond, batch_prompt) 121 | uncond = tensor_to_size(uncond, batch_prompt) 122 | 123 | k_cond = ipadapter.ip_layers.to_kvs[k_key](cond) 124 | k_uncond = ipadapter.ip_layers.to_kvs[k_key](uncond) 125 | v_cond = ipadapter.ip_layers.to_kvs[v_key](cond) 126 | v_uncond = ipadapter.ip_layers.to_kvs[v_key](uncond) 127 | else: 128 | # TODO: should we always convert the weights to a tensor? 129 | if isinstance(weight, torch.Tensor): 130 | weight = tensor_to_size(weight, batch_prompt) 131 | if torch.all(weight == 0): 132 | return 0 133 | weight = weight.repeat(len(cond_or_uncond), 1, 1) # repeat for cond and uncond 134 | elif weight == 0: 135 | return 0 136 | 137 | k_cond = ipadapter.ip_layers.to_kvs[k_key](cond).repeat(batch_prompt, 1, 1) 138 | k_uncond = ipadapter.ip_layers.to_kvs[k_key](uncond).repeat(batch_prompt, 1, 1) 139 | v_cond = ipadapter.ip_layers.to_kvs[v_key](cond).repeat(batch_prompt, 1, 1) 140 | v_uncond = ipadapter.ip_layers.to_kvs[v_key](uncond).repeat(batch_prompt, 1, 1) 141 | 142 | if len(cond_or_uncond) == 3: # TODO: conxl, I need to check this 143 | ip_k = torch.cat([(k_cond, k_uncond, k_cond)[i] for i in cond_or_uncond], dim=0) 144 | ip_v = torch.cat([(v_cond, v_uncond, v_cond)[i] for i in cond_or_uncond], dim=0) 145 | else: 146 | ip_k = torch.cat([(k_cond, k_uncond)[i] for i in cond_or_uncond], dim=0) 147 | ip_v = torch.cat([(v_cond, v_uncond)[i] for i in cond_or_uncond], dim=0) 148 | 149 | if embeds_scaling == 'K+mean(V) w/ C penalty': 150 | scaling = float(ip_k.shape[2]) / 1280.0 151 | weight = weight * scaling 152 | ip_k = ip_k * weight 153 | ip_v_mean = torch.mean(ip_v, dim=1, keepdim=True) 154 | ip_v = (ip_v - ip_v_mean) + ip_v_mean * weight 155 | out_ip = optimized_attention(q, ip_k, ip_v, extra_options["n_heads"]) 156 | del ip_v_mean 157 | elif embeds_scaling == 'K+V w/ C penalty': 158 | scaling = float(ip_k.shape[2]) / 1280.0 159 | weight = weight * scaling 160 | ip_k = ip_k * weight 161 | ip_v = ip_v * weight 162 | out_ip = optimized_attention(q, ip_k, ip_v, extra_options["n_heads"]) 163 | elif embeds_scaling == 'K+V': 164 | ip_k = ip_k * weight 165 | ip_v = ip_v * weight 166 | out_ip = optimized_attention(q, ip_k, ip_v, extra_options["n_heads"]) 167 | else: 168 | #ip_v = ip_v * weight 169 | out_ip = optimized_attention(q, ip_k, ip_v, extra_options["n_heads"]) 170 | out_ip = out_ip * weight # I'm doing this to get the same results as before 171 | 172 | if mask is not None: 173 | mask_h = oh / math.sqrt(oh * ow / seq_len) 174 | mask_h = int(mask_h) + int((seq_len % int(mask_h)) != 0) 175 | mask_w = seq_len // mask_h 176 | 177 | # check if using AnimateDiff and sliding context window 178 | if (mask.shape[0] > 1 and ad_params is not None and ad_params["sub_idxs"] is not None): 179 | # if mask length matches or exceeds full_length, get sub_idx masks 180 | if mask.shape[0] >= ad_params["full_length"]: 181 | mask = torch.Tensor(mask[ad_params["sub_idxs"]]) 182 | mask = F.interpolate(mask.unsqueeze(1), size=(mask_h, mask_w), mode="bilinear").squeeze(1) 183 | else: 184 | mask = F.interpolate(mask.unsqueeze(1), size=(mask_h, mask_w), mode="bilinear").squeeze(1) 185 | mask = tensor_to_size(mask, ad_params["full_length"]) 186 | mask = mask[ad_params["sub_idxs"]] 187 | else: 188 | mask = F.interpolate(mask.unsqueeze(1), size=(mask_h, mask_w), mode="bilinear").squeeze(1) 189 | mask = tensor_to_size(mask, batch_prompt) 190 | 191 | mask = mask.repeat(len(cond_or_uncond), 1, 1) 192 | mask = mask.view(mask.shape[0], -1, 1).repeat(1, 1, out.shape[2]) 193 | 194 | # covers cases where extreme aspect ratios can cause the mask to have a wrong size 195 | mask_len = mask_h * mask_w 196 | if mask_len < seq_len: 197 | pad_len = seq_len - mask_len 198 | pad1 = pad_len // 2 199 | pad2 = pad_len - pad1 200 | mask = F.pad(mask, (0, 0, pad1, pad2), value=0.0) 201 | elif mask_len > seq_len: 202 | crop_start = (mask_len - seq_len) // 2 203 | mask = mask[:, crop_start:crop_start+seq_len, :] 204 | 205 | out_ip = out_ip * mask 206 | 207 | #out = out + out_ip 208 | 209 | return out_ip.to(dtype=dtype) 210 | -------------------------------------------------------------------------------- /ComfyUI_IPAdapter_plus/image_proj_models.py: -------------------------------------------------------------------------------- 1 | import math 2 | import torch 3 | import torch.nn as nn 4 | from einops import rearrange 5 | from einops.layers.torch import Rearrange 6 | 7 | 8 | # FFN 9 | def FeedForward(dim, mult=4): 10 | inner_dim = int(dim * mult) 11 | return nn.Sequential( 12 | nn.LayerNorm(dim), 13 | nn.Linear(dim, inner_dim, bias=False), 14 | nn.GELU(), 15 | nn.Linear(inner_dim, dim, bias=False), 16 | ) 17 | 18 | 19 | def reshape_tensor(x, heads): 20 | bs, length, width = x.shape 21 | # (bs, length, width) --> (bs, length, n_heads, dim_per_head) 22 | x = x.view(bs, length, heads, -1) 23 | # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head) 24 | x = x.transpose(1, 2) 25 | # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head) 26 | x = x.reshape(bs, heads, length, -1) 27 | return x 28 | 29 | 30 | class PerceiverAttention(nn.Module): 31 | def __init__(self, *, dim, dim_head=64, heads=8): 32 | super().__init__() 33 | self.scale = dim_head**-0.5 34 | self.dim_head = dim_head 35 | self.heads = heads 36 | inner_dim = dim_head * heads 37 | 38 | self.norm1 = nn.LayerNorm(dim) 39 | self.norm2 = nn.LayerNorm(dim) 40 | 41 | self.to_q = nn.Linear(dim, inner_dim, bias=False) 42 | self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False) 43 | self.to_out = nn.Linear(inner_dim, dim, bias=False) 44 | 45 | def forward(self, x, latents): 46 | """ 47 | Args: 48 | x (torch.Tensor): image features 49 | shape (b, n1, D) 50 | latent (torch.Tensor): latent features 51 | shape (b, n2, D) 52 | """ 53 | x = self.norm1(x) 54 | latents = self.norm2(latents) 55 | 56 | b, l, _ = latents.shape 57 | 58 | q = self.to_q(latents) 59 | kv_input = torch.cat((x, latents), dim=-2) 60 | k, v = self.to_kv(kv_input).chunk(2, dim=-1) 61 | 62 | q = reshape_tensor(q, self.heads) 63 | k = reshape_tensor(k, self.heads) 64 | v = reshape_tensor(v, self.heads) 65 | 66 | # attention 67 | scale = 1 / math.sqrt(math.sqrt(self.dim_head)) 68 | weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards 69 | weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) 70 | out = weight @ v 71 | 72 | out = out.permute(0, 2, 1, 3).reshape(b, l, -1) 73 | 74 | return self.to_out(out) 75 | 76 | 77 | class Resampler(nn.Module): 78 | def __init__( 79 | self, 80 | dim=1024, 81 | depth=8, 82 | dim_head=64, 83 | heads=16, 84 | num_queries=8, 85 | embedding_dim=768, 86 | output_dim=1024, 87 | ff_mult=4, 88 | max_seq_len: int = 257, # CLIP tokens + CLS token 89 | apply_pos_emb: bool = False, 90 | num_latents_mean_pooled: int = 0, # number of latents derived from mean pooled representation of the sequence 91 | ): 92 | super().__init__() 93 | self.pos_emb = nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None 94 | 95 | self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5) 96 | 97 | self.proj_in = nn.Linear(embedding_dim, dim) 98 | 99 | self.proj_out = nn.Linear(dim, output_dim) 100 | self.norm_out = nn.LayerNorm(output_dim) 101 | 102 | self.to_latents_from_mean_pooled_seq = ( 103 | nn.Sequential( 104 | nn.LayerNorm(dim), 105 | nn.Linear(dim, dim * num_latents_mean_pooled), 106 | Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled), 107 | ) 108 | if num_latents_mean_pooled > 0 109 | else None 110 | ) 111 | 112 | self.layers = nn.ModuleList([]) 113 | for _ in range(depth): 114 | self.layers.append( 115 | nn.ModuleList( 116 | [ 117 | PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads), 118 | FeedForward(dim=dim, mult=ff_mult), 119 | ] 120 | ) 121 | ) 122 | 123 | def forward(self, x): 124 | if self.pos_emb is not None: 125 | n, device = x.shape[1], x.device 126 | pos_emb = self.pos_emb(torch.arange(n, device=device)) 127 | x = x + pos_emb 128 | 129 | latents = self.latents.repeat(x.size(0), 1, 1) 130 | 131 | x = self.proj_in(x) 132 | 133 | if self.to_latents_from_mean_pooled_seq: 134 | meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool)) 135 | meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq) 136 | latents = torch.cat((meanpooled_latents, latents), dim=-2) 137 | 138 | for attn, ff in self.layers: 139 | latents = attn(x, latents) + latents 140 | latents = ff(latents) + latents 141 | 142 | latents = self.proj_out(latents) 143 | return self.norm_out(latents) 144 | 145 | 146 | def masked_mean(t, *, dim, mask=None): 147 | if mask is None: 148 | return t.mean(dim=dim) 149 | 150 | denom = mask.sum(dim=dim, keepdim=True) 151 | mask = rearrange(mask, "b n -> b n 1") 152 | masked_t = t.masked_fill(~mask, 0.0) 153 | 154 | return masked_t.sum(dim=dim) / denom.clamp(min=1e-5) 155 | 156 | 157 | class FacePerceiverResampler(nn.Module): 158 | def __init__( 159 | self, 160 | *, 161 | dim=768, 162 | depth=4, 163 | dim_head=64, 164 | heads=16, 165 | embedding_dim=1280, 166 | output_dim=768, 167 | ff_mult=4, 168 | ): 169 | super().__init__() 170 | 171 | self.proj_in = nn.Linear(embedding_dim, dim) 172 | self.proj_out = nn.Linear(dim, output_dim) 173 | self.norm_out = nn.LayerNorm(output_dim) 174 | self.layers = nn.ModuleList([]) 175 | for _ in range(depth): 176 | self.layers.append( 177 | nn.ModuleList( 178 | [ 179 | PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads), 180 | FeedForward(dim=dim, mult=ff_mult), 181 | ] 182 | ) 183 | ) 184 | 185 | def forward(self, latents, x): 186 | x = self.proj_in(x) 187 | for attn, ff in self.layers: 188 | latents = attn(x, latents) + latents 189 | latents = ff(latents) + latents 190 | latents = self.proj_out(latents) 191 | return self.norm_out(latents) 192 | 193 | 194 | class MLPProjModel(nn.Module): 195 | def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024): 196 | super().__init__() 197 | 198 | self.proj = nn.Sequential( 199 | nn.Linear(clip_embeddings_dim, clip_embeddings_dim), 200 | nn.GELU(), 201 | nn.Linear(clip_embeddings_dim, cross_attention_dim), 202 | nn.LayerNorm(cross_attention_dim) 203 | ) 204 | 205 | def forward(self, image_embeds): 206 | clip_extra_context_tokens = self.proj(image_embeds) 207 | return clip_extra_context_tokens 208 | 209 | class MLPProjModelFaceId(nn.Module): 210 | def __init__(self, cross_attention_dim=768, id_embeddings_dim=512, num_tokens=4): 211 | super().__init__() 212 | 213 | self.cross_attention_dim = cross_attention_dim 214 | self.num_tokens = num_tokens 215 | 216 | self.proj = nn.Sequential( 217 | nn.Linear(id_embeddings_dim, id_embeddings_dim*2), 218 | nn.GELU(), 219 | nn.Linear(id_embeddings_dim*2, cross_attention_dim*num_tokens), 220 | ) 221 | self.norm = nn.LayerNorm(cross_attention_dim) 222 | 223 | def forward(self, id_embeds): 224 | x = self.proj(id_embeds) 225 | x = x.reshape(-1, self.num_tokens, self.cross_attention_dim) 226 | x = self.norm(x) 227 | return x 228 | 229 | class ProjModelFaceIdPlus(nn.Module): 230 | def __init__(self, cross_attention_dim=768, id_embeddings_dim=512, clip_embeddings_dim=1280, num_tokens=4): 231 | super().__init__() 232 | 233 | self.cross_attention_dim = cross_attention_dim 234 | self.num_tokens = num_tokens 235 | 236 | self.proj = nn.Sequential( 237 | nn.Linear(id_embeddings_dim, id_embeddings_dim*2), 238 | nn.GELU(), 239 | nn.Linear(id_embeddings_dim*2, cross_attention_dim*num_tokens), 240 | ) 241 | self.norm = nn.LayerNorm(cross_attention_dim) 242 | 243 | self.perceiver_resampler = FacePerceiverResampler( 244 | dim=cross_attention_dim, 245 | depth=4, 246 | dim_head=64, 247 | heads=cross_attention_dim // 64, 248 | embedding_dim=clip_embeddings_dim, 249 | output_dim=cross_attention_dim, 250 | ff_mult=4, 251 | ) 252 | 253 | def forward(self, id_embeds, clip_embeds, scale=1.0, shortcut=False): 254 | x = self.proj(id_embeds) 255 | x = x.reshape(-1, self.num_tokens, self.cross_attention_dim) 256 | x = self.norm(x) 257 | out = self.perceiver_resampler(x, clip_embeds) 258 | if shortcut: 259 | out = x + scale * out 260 | return out 261 | 262 | class ImageProjModel(nn.Module): 263 | def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4): 264 | super().__init__() 265 | 266 | self.cross_attention_dim = cross_attention_dim 267 | self.clip_extra_context_tokens = clip_extra_context_tokens 268 | self.proj = nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim) 269 | self.norm = nn.LayerNorm(cross_attention_dim) 270 | 271 | def forward(self, image_embeds): 272 | embeds = image_embeds 273 | x = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim) 274 | x = self.norm(x) 275 | return x 276 | -------------------------------------------------------------------------------- /ComfyUI_IPAdapter_plus/utils.py: -------------------------------------------------------------------------------- 1 | import re 2 | import torch 3 | import os 4 | import folder_paths 5 | from comfy.clip_vision import clip_preprocess, Output 6 | import comfy.utils 7 | import comfy.model_management as model_management 8 | try: 9 | import torchvision.transforms.v2 as T 10 | except ImportError: 11 | import torchvision.transforms as T 12 | 13 | def get_clipvision_file(preset): 14 | preset = preset.lower() 15 | clipvision_list = folder_paths.get_filename_list("clip_vision") 16 | 17 | if preset.startswith("vit-g"): 18 | pattern = r'(ViT.bigG.14.*39B.b160k|ipadapter.*sdxl|sdxl.*model\.(bin|safetensors))' 19 | else: 20 | pattern = r'(ViT.H.14.*s32B.b79K|ipadapter.*sd15|sd1.?5.*model\.(bin|safetensors))' 21 | clipvision_file = [e for e in clipvision_list if re.search(pattern, e, re.IGNORECASE)] 22 | 23 | clipvision_file = folder_paths.get_full_path("clip_vision", clipvision_file[0]) if clipvision_file else None 24 | 25 | return clipvision_file 26 | 27 | def get_ipadapter_file(preset, is_sdxl): 28 | preset = preset.lower() 29 | ipadapter_list = folder_paths.get_filename_list("ipadapter") 30 | is_insightface = False 31 | lora_pattern = None 32 | 33 | if preset.startswith("light"): 34 | if is_sdxl: 35 | raise Exception("light model is not supported for SDXL") 36 | pattern = r'sd15.light.v11\.(safetensors|bin)$' 37 | # if v11 is not found, try with the old version 38 | if not [e for e in ipadapter_list if re.search(pattern, e, re.IGNORECASE)]: 39 | pattern = r'sd15.light\.(safetensors|bin)$' 40 | elif preset.startswith("standard"): 41 | if is_sdxl: 42 | pattern = r'ip.adapter.sdxl.vit.h\.(safetensors|bin)$' 43 | else: 44 | pattern = r'ip.adapter.sd15\.(safetensors|bin)$' 45 | elif preset.startswith("vit-g"): 46 | if is_sdxl: 47 | pattern = r'ip.adapter.sdxl\.(safetensors|bin)$' 48 | else: 49 | pattern = r'sd15.vit.g\.(safetensors|bin)$' 50 | elif preset.startswith("plus ("): 51 | if is_sdxl: 52 | pattern = r'plus.sdxl.vit.h\.(safetensors|bin)$' 53 | else: 54 | pattern = r'ip.adapter.plus.sd15\.(safetensors|bin)$' 55 | elif preset.startswith("plus face"): 56 | if is_sdxl: 57 | pattern = r'plus.face.sdxl.vit.h\.(safetensors|bin)$' 58 | else: 59 | pattern = r'plus.face.sd15\.(safetensors|bin)$' 60 | elif preset.startswith("full"): 61 | if is_sdxl: 62 | raise Exception("full face model is not supported for SDXL") 63 | pattern = r'full.face.sd15\.(safetensors|bin)$' 64 | elif preset.startswith("faceid portrait ("): 65 | if is_sdxl: 66 | pattern = r'portrait.sdxl\.(safetensors|bin)$' 67 | else: 68 | pattern = r'portrait.v11.sd15\.(safetensors|bin)$' 69 | # if v11 is not found, try with the old version 70 | if not [e for e in ipadapter_list if re.search(pattern, e, re.IGNORECASE)]: 71 | pattern = r'portrait.sd15\.(safetensors|bin)$' 72 | is_insightface = True 73 | elif preset.startswith("faceid portrait unnorm"): 74 | if is_sdxl: 75 | pattern = r'portrait.sdxl.unnorm\.(safetensors|bin)$' 76 | else: 77 | raise Exception("portrait unnorm model is not supported for SD1.5") 78 | is_insightface = True 79 | elif preset == "faceid": 80 | if is_sdxl: 81 | pattern = r'faceid.sdxl\.(safetensors|bin)$' 82 | lora_pattern = r'faceid.sdxl.lora\.safetensors$' 83 | else: 84 | pattern = r'faceid.sd15\.(safetensors|bin)$' 85 | lora_pattern = r'faceid.sd15.lora\.safetensors$' 86 | is_insightface = True 87 | elif preset.startswith("faceid plus -"): 88 | if is_sdxl: 89 | raise Exception("faceid plus model is not supported for SDXL") 90 | pattern = r'faceid.plus.sd15\.(safetensors|bin)$' 91 | lora_pattern = r'faceid.plus.sd15.lora\.safetensors$' 92 | is_insightface = True 93 | elif preset.startswith("faceid plus v2"): 94 | if is_sdxl: 95 | pattern = r'faceid.plusv2.sdxl\.(safetensors|bin)$' 96 | lora_pattern = r'faceid.plusv2.sdxl.lora\.safetensors$' 97 | else: 98 | pattern = r'faceid.plusv2.sd15\.(safetensors|bin)$' 99 | lora_pattern = r'faceid.plusv2.sd15.lora\.safetensors$' 100 | is_insightface = True 101 | # Community's models 102 | elif preset.startswith("composition"): 103 | if is_sdxl: 104 | pattern = r'plus.composition.sdxl\.safetensors$' 105 | else: 106 | pattern = r'plus.composition.sd15\.safetensors$' 107 | else: 108 | raise Exception(f"invalid type '{preset}'") 109 | 110 | ipadapter_file = [e for e in ipadapter_list if re.search(pattern, e, re.IGNORECASE)] 111 | ipadapter_file = folder_paths.get_full_path("ipadapter", ipadapter_file[0]) if ipadapter_file else None 112 | 113 | return ipadapter_file, is_insightface, lora_pattern 114 | 115 | def get_lora_file(pattern): 116 | lora_list = folder_paths.get_filename_list("loras") 117 | lora_file = [e for e in lora_list if re.search(pattern, e, re.IGNORECASE)] 118 | lora_file = folder_paths.get_full_path("loras", lora_file[0]) if lora_file else None 119 | 120 | return lora_file 121 | 122 | def ipadapter_model_loader(file): 123 | model = comfy.utils.load_torch_file(file, safe_load=True) 124 | 125 | if file.lower().endswith(".safetensors"): 126 | st_model = {"image_proj": {}, "ip_adapter": {}} 127 | for key in model.keys(): 128 | if key.startswith("image_proj."): 129 | st_model["image_proj"][key.replace("image_proj.", "")] = model[key] 130 | elif key.startswith("ip_adapter."): 131 | st_model["ip_adapter"][key.replace("ip_adapter.", "")] = model[key] 132 | model = st_model 133 | del st_model 134 | 135 | if "adapter_modules" in model.keys(): 136 | model["ip_adapter"] = model["adapter_modules"] 137 | del model["adapter_modules"] 138 | 139 | if not "ip_adapter" in model.keys() or not model["ip_adapter"]: 140 | raise Exception("invalid IPAdapter model {}".format(file)) 141 | 142 | if 'plusv2' in file.lower(): 143 | model["faceidplusv2"] = True 144 | 145 | if 'unnorm' in file.lower(): 146 | model["portraitunnorm"] = True 147 | 148 | return model 149 | 150 | def insightface_loader(provider): 151 | try: 152 | from insightface.app import FaceAnalysis 153 | except ImportError as e: 154 | raise Exception(e) 155 | 156 | path = os.path.join(folder_paths.models_dir, "insightface") 157 | model = FaceAnalysis(name="antelopev2", root=path, providers=[provider + 'ExecutionProvider',]) 158 | model.prepare(ctx_id=0, det_size=(640, 640)) 159 | return model 160 | 161 | def encode_image_masked(clip_vision, image, mask=None, batch_size=0, size=224): 162 | model_management.load_model_gpu(clip_vision.patcher) 163 | outputs = Output() 164 | 165 | if batch_size == 0: 166 | batch_size = image.shape[0] 167 | elif batch_size > image.shape[0]: 168 | batch_size = image.shape[0] 169 | 170 | image_batch = torch.split(image, batch_size, dim=0) 171 | 172 | for img in image_batch: 173 | img = img.to(clip_vision.load_device) 174 | 175 | pixel_values = clip_preprocess(img.to(clip_vision.load_device), size=size).float() 176 | 177 | # TODO: support for multiple masks 178 | if mask is not None: 179 | pixel_values = pixel_values * mask.to(clip_vision.load_device) 180 | 181 | out = clip_vision.model(pixel_values=pixel_values, intermediate_output=-2) 182 | 183 | if not hasattr(outputs, "last_hidden_state"): 184 | outputs["last_hidden_state"] = out[0].to(model_management.intermediate_device()) 185 | outputs["image_embeds"] = out[2].to(model_management.intermediate_device()) 186 | outputs["penultimate_hidden_states"] = out[1].to(model_management.intermediate_device()) 187 | else: 188 | outputs["last_hidden_state"] = torch.cat((outputs["last_hidden_state"], out[0].to(model_management.intermediate_device())), dim=0) 189 | outputs["image_embeds"] = torch.cat((outputs["image_embeds"], out[2].to(model_management.intermediate_device())), dim=0) 190 | outputs["penultimate_hidden_states"] = torch.cat((outputs["penultimate_hidden_states"], out[1].to(model_management.intermediate_device())), dim=0) 191 | 192 | del img, pixel_values, out 193 | torch.cuda.empty_cache() 194 | 195 | return outputs 196 | 197 | def tensor_to_size(source, dest_size): 198 | if isinstance(dest_size, torch.Tensor): 199 | dest_size = dest_size.shape[0] 200 | source_size = source.shape[0] 201 | 202 | if source_size < dest_size: 203 | shape = [dest_size - source_size] + [1]*(source.dim()-1) 204 | source = torch.cat((source, source[-1:].repeat(shape)), dim=0) 205 | elif source_size > dest_size: 206 | source = source[:dest_size] 207 | 208 | return source 209 | 210 | def min_(tensor_list): 211 | # return the element-wise min of the tensor list. 212 | x = torch.stack(tensor_list) 213 | mn = x.min(axis=0)[0] 214 | return torch.clamp(mn, min=0) 215 | 216 | def max_(tensor_list): 217 | # return the element-wise max of the tensor list. 218 | x = torch.stack(tensor_list) 219 | mx = x.max(axis=0)[0] 220 | return torch.clamp(mx, max=1) 221 | 222 | # From https://github.com/Jamy-L/Pytorch-Contrast-Adaptive-Sharpening/ 223 | def contrast_adaptive_sharpening(image, amount): 224 | img = T.functional.pad(image, (1, 1, 1, 1)).cpu() 225 | 226 | a = img[..., :-2, :-2] 227 | b = img[..., :-2, 1:-1] 228 | c = img[..., :-2, 2:] 229 | d = img[..., 1:-1, :-2] 230 | e = img[..., 1:-1, 1:-1] 231 | f = img[..., 1:-1, 2:] 232 | g = img[..., 2:, :-2] 233 | h = img[..., 2:, 1:-1] 234 | i = img[..., 2:, 2:] 235 | 236 | # Computing contrast 237 | cross = (b, d, e, f, h) 238 | mn = min_(cross) 239 | mx = max_(cross) 240 | 241 | diag = (a, c, g, i) 242 | mn2 = min_(diag) 243 | mx2 = max_(diag) 244 | mx = mx + mx2 245 | mn = mn + mn2 246 | 247 | # Computing local weight 248 | inv_mx = torch.reciprocal(mx) 249 | amp = inv_mx * torch.minimum(mn, (2 - mx)) 250 | 251 | # scaling 252 | amp = torch.sqrt(amp) 253 | w = - amp * (amount * (1/5 - 1/8) + 1/8) 254 | div = torch.reciprocal(1 + 4*w) 255 | 256 | output = ((b + d + f + h)*w + e) * div 257 | output = torch.nan_to_num(output) 258 | output = output.clamp(0, 1) 259 | 260 | return output 261 | 262 | def tensor_to_image(tensor): 263 | image = tensor.mul(255).clamp(0, 255).byte().cpu() 264 | image = image[..., [2, 1, 0]].numpy() 265 | return image 266 | 267 | def image_to_tensor(image): 268 | tensor = torch.clamp(torch.from_numpy(image).float() / 255., 0, 1) 269 | tensor = tensor[..., [2, 1, 0]] 270 | return tensor 271 | -------------------------------------------------------------------------------- /__init__.py: -------------------------------------------------------------------------------- 1 | import inspect 2 | import json 3 | import os 4 | import folder_paths 5 | import importlib 6 | 7 | 8 | NODE_CLASS_MAPPINGS = { 9 | } 10 | 11 | 12 | NODE_DISPLAY_NAME_MAPPINGS = { 13 | } 14 | 15 | MAX_RESOLUTION = 16384 16 | 17 | AUTHOR_NAME = "MinusZone" 18 | CATEGORY_NAME = f"{AUTHOR_NAME} - Kolors" 19 | folder_paths.add_model_folder_path( 20 | "LLM", os.path.join(folder_paths.models_dir, "LLM")) 21 | 22 | 23 | class MZ_ChatGLM3Loader: 24 | @classmethod 25 | def INPUT_TYPES(s): 26 | # from .mz_kolors_utils import Utils 27 | # llm_dir = os.path.join(Utils.get_models_path(), "LLM") 28 | # print("llm_dir:", llm_dir) 29 | llm_models = folder_paths.get_filename_list("LLM") 30 | 31 | # 筛选safetensors结尾的文件 32 | llm_models = [ 33 | model for model in llm_models if model.endswith("safetensors")] 34 | 35 | return {"required": { 36 | "chatglm3_checkpoint": (llm_models,), 37 | }} 38 | 39 | RETURN_TYPES = ("CHATGLM3MODEL",) 40 | RETURN_NAMES = ("chatglm3_model",) 41 | FUNCTION = "load_chatglm3" 42 | CATEGORY = CATEGORY_NAME 43 | 44 | def load_chatglm3(self, **kwargs): 45 | from . import mz_kolors_core 46 | importlib.reload(mz_kolors_core) 47 | return mz_kolors_core.MZ_ChatGLM3Loader_call(kwargs) 48 | 49 | 50 | NODE_CLASS_MAPPINGS["MZ_ChatGLM3Loader"] = MZ_ChatGLM3Loader 51 | NODE_DISPLAY_NAME_MAPPINGS["MZ_ChatGLM3Loader"] = f"{AUTHOR_NAME} - ChatGLM3Loader" 52 | 53 | 54 | class MZ_ChatGLM3TextEncodeV2: 55 | @classmethod 56 | def INPUT_TYPES(s): 57 | return { 58 | "required": { 59 | "chatglm3_model": ("CHATGLM3MODEL", ), 60 | "text": ("STRING", {"multiline": True, "dynamicPrompts": True}), 61 | } 62 | } 63 | 64 | RETURN_TYPES = ("CONDITIONING",) 65 | 66 | FUNCTION = "encode" 67 | CATEGORY = CATEGORY_NAME 68 | 69 | def encode(self, **kwargs): 70 | from . import mz_kolors_core 71 | importlib.reload(mz_kolors_core) 72 | return mz_kolors_core.MZ_ChatGLM3TextEncodeV2_call(kwargs) 73 | 74 | 75 | NODE_CLASS_MAPPINGS["MZ_ChatGLM3_V2"] = MZ_ChatGLM3TextEncodeV2 76 | NODE_DISPLAY_NAME_MAPPINGS[ 77 | "MZ_ChatGLM3_V2"] = f"{AUTHOR_NAME} - ChatGLM3TextEncodeV2" 78 | 79 | 80 | class MZ_ChatGLM3Embeds2Conditioning: 81 | @classmethod 82 | def INPUT_TYPES(s): 83 | return { 84 | "required": { 85 | "kolors_embeds": ("KOLORS_EMBEDS", ), 86 | "width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), 87 | "height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), 88 | "crop_w": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION}), 89 | "crop_h": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION}), 90 | "target_width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), 91 | "target_height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), 92 | } 93 | } 94 | 95 | RETURN_TYPES = ("CONDITIONING", "CONDITIONING",) 96 | RETURN_NAMES = ("positive", "negative",) 97 | 98 | FUNCTION = "embeds2conditioning" 99 | CATEGORY = CATEGORY_NAME 100 | 101 | def embeds2conditioning(self, **kwargs): 102 | from . import mz_kolors_core 103 | importlib.reload(mz_kolors_core) 104 | return mz_kolors_core.MZ_ChatGLM3Embeds2Conditioning_call(kwargs) 105 | 106 | 107 | NODE_CLASS_MAPPINGS["MZ_ChatGLM3Embeds2Conditioning"] = MZ_ChatGLM3Embeds2Conditioning 108 | NODE_DISPLAY_NAME_MAPPINGS[ 109 | "MZ_ChatGLM3Embeds2Conditioning"] = f"{AUTHOR_NAME} - ChatGLM3Embeds2Conditioning" 110 | 111 | 112 | # for 2048 resolution 113 | class MZ_ChatGLM3TextEncodeAdvanceV2: 114 | @classmethod 115 | def INPUT_TYPES(s): 116 | return { 117 | "required": { 118 | "chatglm3_model": ("CHATGLM3MODEL", ), 119 | "text": ("STRING", {"multiline": True, "dynamicPrompts": True}), 120 | "width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), 121 | "height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), 122 | "crop_w": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION}), 123 | "crop_h": ("INT", {"default": 0, "min": 0, "max": MAX_RESOLUTION}), 124 | "target_width": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), 125 | "target_height": ("INT", {"default": 1024.0, "min": 0, "max": MAX_RESOLUTION}), 126 | } 127 | } 128 | 129 | RETURN_TYPES = ("CONDITIONING",) 130 | 131 | FUNCTION = "encode" 132 | CATEGORY = CATEGORY_NAME 133 | 134 | def encode(self, **kwargs): 135 | from . import mz_kolors_core 136 | importlib.reload(mz_kolors_core) 137 | return mz_kolors_core.MZ_ChatGLM3TextEncodeV2_call(kwargs) 138 | 139 | 140 | NODE_CLASS_MAPPINGS["MZ_ChatGLM3_Advance_V2"] = MZ_ChatGLM3TextEncodeAdvanceV2 141 | NODE_DISPLAY_NAME_MAPPINGS[ 142 | "MZ_ChatGLM3_Advance_V2"] = f"{AUTHOR_NAME} - ChatGLM3TextEncodeAdvanceV2" 143 | 144 | 145 | class MZ_KolorsCheckpointLoaderSimple(): 146 | @classmethod 147 | def INPUT_TYPES(s): 148 | return {"required": {"ckpt_name": (folder_paths.get_filename_list("checkpoints"), ), 149 | }} 150 | RETURN_TYPES = ("MODEL", "VAE") 151 | FUNCTION = "load_checkpoint" 152 | 153 | CATEGORY = CATEGORY_NAME 154 | 155 | def load_checkpoint(self, **kwargs): 156 | from . import mz_kolors_core 157 | importlib.reload(mz_kolors_core) 158 | return mz_kolors_core.MZ_KolorsCheckpointLoaderSimple_call(kwargs) 159 | 160 | 161 | NODE_CLASS_MAPPINGS["MZ_KolorsCheckpointLoaderSimple"] = MZ_KolorsCheckpointLoaderSimple 162 | NODE_DISPLAY_NAME_MAPPINGS[ 163 | "MZ_KolorsCheckpointLoaderSimple"] = f"{AUTHOR_NAME} - KolorsCheckpointLoaderSimple" 164 | 165 | 166 | class MZ_KolorsControlNetLoader: 167 | @classmethod 168 | def INPUT_TYPES(s): 169 | return {"required": { 170 | "control_net_name": (folder_paths.get_filename_list("controlnet"), ), 171 | # "seed": ("INT", {"default": 0, "min": 0, "max": 1000000}), 172 | }} 173 | 174 | RETURN_TYPES = ("CONTROL_NET",) 175 | RETURN_NAMES = ("ControlNet",) 176 | FUNCTION = "load_controlnet" 177 | 178 | CATEGORY = CATEGORY_NAME 179 | 180 | def load_controlnet(self, **kwargs): 181 | from . import mz_kolors_core 182 | importlib.reload(mz_kolors_core) 183 | return mz_kolors_core.MZ_KolorsControlNetLoader_call(kwargs) 184 | 185 | 186 | NODE_CLASS_MAPPINGS["MZ_KolorsControlNetLoader"] = MZ_KolorsControlNetLoader 187 | NODE_DISPLAY_NAME_MAPPINGS[ 188 | "MZ_KolorsControlNetLoader"] = f"{AUTHOR_NAME} - KolorsControlNetLoader" 189 | 190 | 191 | class MZ_KolorsUNETLoaderV2(): 192 | @classmethod 193 | def INPUT_TYPES(s): 194 | return {"required": { 195 | "unet_name": (folder_paths.get_filename_list("unet"), ), 196 | }} 197 | 198 | RETURN_TYPES = ("MODEL",) 199 | 200 | RETURN_NAMES = ("model",) 201 | 202 | FUNCTION = "load_unet" 203 | 204 | CATEGORY = CATEGORY_NAME 205 | 206 | def load_unet(self, **kwargs): 207 | from . import mz_kolors_core 208 | importlib.reload(mz_kolors_core) 209 | return mz_kolors_core.MZ_KolorsUNETLoaderV2_call(kwargs) 210 | 211 | 212 | NODE_CLASS_MAPPINGS["MZ_KolorsUNETLoaderV2"] = MZ_KolorsUNETLoaderV2 213 | NODE_DISPLAY_NAME_MAPPINGS[ 214 | "MZ_KolorsUNETLoaderV2"] = f"{AUTHOR_NAME} - KolorsUNETLoaderV2" 215 | 216 | 217 | class MZ_KolorsControlNetPatch: 218 | @classmethod 219 | def INPUT_TYPES(s): 220 | return { 221 | "required": { 222 | "control_net": ("CONTROL_NET", ), 223 | "model": ("MODEL", ), 224 | } 225 | } 226 | 227 | RETURN_TYPES = ("CONTROL_NET",) 228 | 229 | FUNCTION = "start" 230 | CATEGORY = CATEGORY_NAME 231 | 232 | def start(self, **kwargs): 233 | from . import mz_kolors_core 234 | importlib.reload(mz_kolors_core) 235 | return mz_kolors_core.MZ_KolorsControlNetPatch_call(kwargs) 236 | 237 | 238 | NODE_CLASS_MAPPINGS["MZ_KolorsControlNetPatch"] = MZ_KolorsControlNetPatch 239 | NODE_DISPLAY_NAME_MAPPINGS[ 240 | "MZ_KolorsControlNetPatch"] = f"{AUTHOR_NAME} - KolorsControlNetPatch" 241 | 242 | 243 | class MZ_KolorsCLIPVisionLoader: 244 | @classmethod 245 | def INPUT_TYPES(s): 246 | return {"required": {"clip_name": (folder_paths.get_filename_list("clip_vision"), ), 247 | }} 248 | RETURN_TYPES = ("CLIP_VISION",) 249 | FUNCTION = "load_clip" 250 | 251 | CATEGORY = CATEGORY_NAME + "/Legacy" 252 | 253 | def load_clip(self, **kwargs): 254 | from . import mz_kolors_core 255 | importlib.reload(mz_kolors_core) 256 | return mz_kolors_core.MZ_KolorsCLIPVisionLoader_call(kwargs) 257 | 258 | 259 | NODE_CLASS_MAPPINGS["MZ_KolorsCLIPVisionLoader"] = MZ_KolorsCLIPVisionLoader 260 | NODE_DISPLAY_NAME_MAPPINGS["MZ_KolorsCLIPVisionLoader"] = f"{AUTHOR_NAME} - KolorsCLIPVisionLoader - Legacy" 261 | 262 | 263 | class MZ_ApplySDXLSamplingSettings(): 264 | @classmethod 265 | def INPUT_TYPES(s): 266 | return { 267 | "required": { 268 | "model": ("MODEL", ), 269 | } 270 | } 271 | 272 | RETURN_TYPES = ("MODEL", ) 273 | 274 | FUNCTION = "apply_sampling_settings" 275 | CATEGORY = CATEGORY_NAME 276 | 277 | def apply_sampling_settings(self, **kwargs): 278 | from . import mz_kolors_core 279 | importlib.reload(mz_kolors_core) 280 | return mz_kolors_core.MZ_ApplySDXLSamplingSettings_call(kwargs) 281 | 282 | 283 | NODE_CLASS_MAPPINGS["MZ_ApplySDXLSamplingSettings"] = MZ_ApplySDXLSamplingSettings 284 | NODE_DISPLAY_NAME_MAPPINGS[ 285 | "MZ_ApplySDXLSamplingSettings"] = f"{AUTHOR_NAME} - ApplySDXLSamplingSettings" 286 | 287 | 288 | class MZ_ApplyCUDAGenerator(): 289 | @classmethod 290 | def INPUT_TYPES(s): 291 | return { 292 | "required": { 293 | "model": ("MODEL", ), 294 | } 295 | } 296 | 297 | RETURN_TYPES = ("MODEL", ) 298 | 299 | FUNCTION = "apply_cuda_generator" 300 | CATEGORY = CATEGORY_NAME 301 | 302 | def apply_cuda_generator(self, **kwargs): 303 | from . import mz_kolors_core 304 | importlib.reload(mz_kolors_core) 305 | return mz_kolors_core.MZ_ApplyCUDAGenerator_call(kwargs) 306 | 307 | 308 | NODE_CLASS_MAPPINGS["MZ_ApplyCUDAGenerator"] = MZ_ApplyCUDAGenerator 309 | NODE_DISPLAY_NAME_MAPPINGS[ 310 | "MZ_ApplyCUDAGenerator"] = f"{AUTHOR_NAME} - ApplyCUDAGenerator" 311 | 312 | 313 | from .ComfyUI_IPAdapter_plus.IPAdapterPlus import IPAdapterAdvanced, IPAdapterModelLoader, IPAdapterInsightFaceLoader, IPAdapterFaceID 314 | 315 | IPAdapterModelLoader.CATEGORY = CATEGORY_NAME + "/IPAdapter" 316 | NODE_CLASS_MAPPINGS["MZ_IPAdapterModelLoaderKolors"] = IPAdapterModelLoader 317 | NODE_DISPLAY_NAME_MAPPINGS[ 318 | "MZ_IPAdapterModelLoaderKolors"] = f"IPAdapterModelLoader(kolors) - Legacy" 319 | 320 | IPAdapterAdvanced.CATEGORY = CATEGORY_NAME + "/IPAdapter" 321 | NODE_CLASS_MAPPINGS["MZ_IPAdapterAdvancedKolors"] = IPAdapterAdvanced 322 | NODE_DISPLAY_NAME_MAPPINGS[ 323 | "MZ_IPAdapterAdvancedKolors"] = f"IPAdapterAdvanced(kolors) - Legacy" 324 | 325 | IPAdapterInsightFaceLoader.CATEGORY = CATEGORY_NAME + "/IPAdapter" 326 | NODE_CLASS_MAPPINGS["MZ_IPAdapterInsightFaceLoader"] = IPAdapterInsightFaceLoader 327 | 328 | NODE_DISPLAY_NAME_MAPPINGS[ 329 | "MZ_IPAdapterInsightFaceLoader"] = f"IPAdapterInsightFaceLoader(kolors) - Legacy" 330 | 331 | IPAdapterFaceID.CATEGORY = CATEGORY_NAME + "/IPAdapter" 332 | NODE_CLASS_MAPPINGS["MZ_IPAdapterFaceID"] = IPAdapterFaceID 333 | 334 | NODE_DISPLAY_NAME_MAPPINGS[ 335 | "MZ_IPAdapterFaceID"] = f"IPAdapterFaceID(kolors) - Legacy" 336 | 337 | from . import mz_kolors_legacy 338 | NODE_CLASS_MAPPINGS.update(mz_kolors_legacy.NODE_CLASS_MAPPINGS) 339 | NODE_DISPLAY_NAME_MAPPINGS.update(mz_kolors_legacy.NODE_DISPLAY_NAME_MAPPINGS) 340 | -------------------------------------------------------------------------------- /chatglm3/tokenization_chatglm.py: -------------------------------------------------------------------------------- 1 | import json 2 | import os 3 | import re 4 | from typing import List, Optional, Union, Dict 5 | from sentencepiece import SentencePieceProcessor 6 | from transformers import PreTrainedTokenizer 7 | from transformers.utils import logging, PaddingStrategy 8 | from transformers.tokenization_utils_base import EncodedInput, BatchEncoding 9 | 10 | 11 | class SPTokenizer: 12 | def __init__(self, model_path: str): 13 | # reload tokenizer 14 | assert os.path.isfile(model_path), model_path 15 | self.sp_model = SentencePieceProcessor(model_file=model_path) 16 | 17 | # BOS / EOS token IDs 18 | self.n_words: int = self.sp_model.vocab_size() 19 | self.bos_id: int = self.sp_model.bos_id() 20 | self.eos_id: int = self.sp_model.eos_id() 21 | self.pad_id: int = self.sp_model.unk_id() 22 | assert self.sp_model.vocab_size() == self.sp_model.get_piece_size() 23 | 24 | role_special_tokens = ["<|system|>", "<|user|>", 25 | "<|assistant|>", "<|observation|>"] 26 | special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", 27 | "sop", "eop"] + role_special_tokens 28 | self.special_tokens = {} 29 | self.index_special_tokens = {} 30 | for token in special_tokens: 31 | self.special_tokens[token] = self.n_words 32 | self.index_special_tokens[self.n_words] = token 33 | self.n_words += 1 34 | self.role_special_token_expression = "|".join( 35 | [re.escape(token) for token in role_special_tokens]) 36 | 37 | def tokenize(self, s: str, encode_special_tokens=False): 38 | if encode_special_tokens: 39 | last_index = 0 40 | t = [] 41 | for match in re.finditer(self.role_special_token_expression, s): 42 | if last_index < match.start(): 43 | t.extend(self.sp_model.EncodeAsPieces( 44 | s[last_index:match.start()])) 45 | t.append(s[match.start():match.end()]) 46 | last_index = match.end() 47 | if last_index < len(s): 48 | t.extend(self.sp_model.EncodeAsPieces(s[last_index:])) 49 | return t 50 | else: 51 | return self.sp_model.EncodeAsPieces(s) 52 | 53 | def encode(self, s: str, bos: bool = False, eos: bool = False) -> List[int]: 54 | assert type(s) is str 55 | t = self.sp_model.encode(s) 56 | if bos: 57 | t = [self.bos_id] + t 58 | if eos: 59 | t = t + [self.eos_id] 60 | return t 61 | 62 | def decode(self, t: List[int]) -> str: 63 | text, buffer = "", [] 64 | for token in t: 65 | if token in self.index_special_tokens: 66 | if buffer: 67 | text += self.sp_model.decode(buffer) 68 | buffer = [] 69 | text += self.index_special_tokens[token] 70 | else: 71 | buffer.append(token) 72 | if buffer: 73 | text += self.sp_model.decode(buffer) 74 | return text 75 | 76 | def decode_tokens(self, tokens: List[str]) -> str: 77 | text = self.sp_model.DecodePieces(tokens) 78 | return text 79 | 80 | def convert_token_to_id(self, token): 81 | """ Converts a token (str) in an id using the vocab. """ 82 | if token in self.special_tokens: 83 | return self.special_tokens[token] 84 | return self.sp_model.PieceToId(token) 85 | 86 | def convert_id_to_token(self, index): 87 | """Converts an index (integer) in a token (str) using the vocab.""" 88 | if index in self.index_special_tokens: 89 | return self.index_special_tokens[index] 90 | if index in [self.eos_id, self.bos_id, self.pad_id] or index < 0: 91 | return "" 92 | return self.sp_model.IdToPiece(index) 93 | 94 | 95 | class ChatGLMTokenizer(PreTrainedTokenizer): 96 | vocab_files_names = {"vocab_file": "tokenizer.model"} 97 | 98 | model_input_names = ["input_ids", "attention_mask", "position_ids"] 99 | 100 | def __init__(self, vocab_file, padding_side="left", clean_up_tokenization_spaces=False, encode_special_tokens=False, 101 | **kwargs): 102 | self.name = "GLMTokenizer" 103 | 104 | self.vocab_file = vocab_file 105 | self.tokenizer = SPTokenizer(vocab_file) 106 | self.special_tokens = { 107 | "": self.tokenizer.bos_id, 108 | "": self.tokenizer.eos_id, 109 | "": self.tokenizer.pad_id 110 | } 111 | self.encode_special_tokens = encode_special_tokens 112 | super().__init__(padding_side=padding_side, clean_up_tokenization_spaces=clean_up_tokenization_spaces, 113 | encode_special_tokens=encode_special_tokens, 114 | **kwargs) 115 | 116 | def get_command(self, token): 117 | if token in self.special_tokens: 118 | return self.special_tokens[token] 119 | assert token in self.tokenizer.special_tokens, f"{token} is not a special token for {self.name}" 120 | return self.tokenizer.special_tokens[token] 121 | 122 | @property 123 | def unk_token(self) -> str: 124 | return "" 125 | 126 | @property 127 | def pad_token(self) -> str: 128 | return "" 129 | 130 | @property 131 | def pad_token_id(self): 132 | return self.get_command("") 133 | 134 | @property 135 | def eos_token(self) -> str: 136 | return "" 137 | 138 | @property 139 | def eos_token_id(self): 140 | return self.get_command("") 141 | 142 | @property 143 | def vocab_size(self): 144 | return self.tokenizer.n_words 145 | 146 | def get_vocab(self): 147 | """ Returns vocab as a dict """ 148 | vocab = {self._convert_id_to_token( 149 | i): i for i in range(self.vocab_size)} 150 | vocab.update(self.added_tokens_encoder) 151 | return vocab 152 | 153 | def _tokenize(self, text, **kwargs): 154 | return self.tokenizer.tokenize(text, encode_special_tokens=self.encode_special_tokens) 155 | 156 | def _convert_token_to_id(self, token): 157 | """ Converts a token (str) in an id using the vocab. """ 158 | return self.tokenizer.convert_token_to_id(token) 159 | 160 | def _convert_id_to_token(self, index): 161 | """Converts an index (integer) in a token (str) using the vocab.""" 162 | return self.tokenizer.convert_id_to_token(index) 163 | 164 | def convert_tokens_to_string(self, tokens: List[str]) -> str: 165 | return self.tokenizer.decode_tokens(tokens) 166 | 167 | def save_vocabulary(self, save_directory, filename_prefix=None): 168 | """ 169 | Save the vocabulary and special tokens file to a directory. 170 | 171 | Args: 172 | save_directory (`str`): 173 | The directory in which to save the vocabulary. 174 | filename_prefix (`str`, *optional*): 175 | An optional prefix to add to the named of the saved files. 176 | 177 | Returns: 178 | `Tuple(str)`: Paths to the files saved. 179 | """ 180 | if os.path.isdir(save_directory): 181 | vocab_file = os.path.join( 182 | save_directory, self.vocab_files_names["vocab_file"] 183 | ) 184 | else: 185 | vocab_file = save_directory 186 | 187 | with open(self.vocab_file, 'rb') as fin: 188 | proto_str = fin.read() 189 | 190 | with open(vocab_file, "wb") as writer: 191 | writer.write(proto_str) 192 | 193 | return (vocab_file,) 194 | 195 | def get_prefix_tokens(self): 196 | prefix_tokens = [self.get_command("[gMASK]"), self.get_command("sop")] 197 | return prefix_tokens 198 | 199 | def build_single_message(self, role, metadata, message): 200 | assert role in ["system", "user", "assistant", "observation"], role 201 | role_tokens = [self.get_command( 202 | f"<|{role}|>")] + self.tokenizer.encode(f"{metadata}\n") 203 | message_tokens = self.tokenizer.encode(message) 204 | tokens = role_tokens + message_tokens 205 | return tokens 206 | 207 | def build_chat_input(self, query, history=None, role="user"): 208 | if history is None: 209 | history = [] 210 | input_ids = [] 211 | for item in history: 212 | content = item["content"] 213 | if item["role"] == "system" and "tools" in item: 214 | content = content + "\n" + \ 215 | json.dumps(item["tools"], indent=4, ensure_ascii=False) 216 | input_ids.extend(self.build_single_message( 217 | item["role"], item.get("metadata", ""), content)) 218 | input_ids.extend(self.build_single_message(role, "", query)) 219 | input_ids.extend([self.get_command("<|assistant|>")]) 220 | return self.batch_encode_plus([input_ids], return_tensors="pt", is_split_into_words=True) 221 | 222 | def build_inputs_with_special_tokens( 223 | self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None 224 | ) -> List[int]: 225 | """ 226 | Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and 227 | adding special tokens. A BERT sequence has the following format: 228 | 229 | - single sequence: `[CLS] X [SEP]` 230 | - pair of sequences: `[CLS] A [SEP] B [SEP]` 231 | 232 | Args: 233 | token_ids_0 (`List[int]`): 234 | List of IDs to which the special tokens will be added. 235 | token_ids_1 (`List[int]`, *optional*): 236 | Optional second list of IDs for sequence pairs. 237 | 238 | Returns: 239 | `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens. 240 | """ 241 | prefix_tokens = self.get_prefix_tokens() 242 | token_ids_0 = prefix_tokens + token_ids_0 243 | if token_ids_1 is not None: 244 | token_ids_0 = token_ids_0 + token_ids_1 + \ 245 | [self.get_command("")] 246 | return token_ids_0 247 | 248 | def _pad( 249 | self, 250 | encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding], 251 | max_length: Optional[int] = None, 252 | padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD, 253 | pad_to_multiple_of: Optional[int] = None, 254 | return_attention_mask: Optional[bool] = None, 255 | **kwargs, 256 | ) -> dict: 257 | """ 258 | Pad encoded inputs (on left/right and up to predefined length or max length in the batch) 259 | 260 | Args: 261 | encoded_inputs: 262 | Dictionary of tokenized inputs (`List[int]`) or batch of tokenized inputs (`List[List[int]]`). 263 | max_length: maximum length of the returned list and optionally padding length (see below). 264 | Will truncate by taking into account the special tokens. 265 | padding_strategy: PaddingStrategy to use for padding. 266 | 267 | - PaddingStrategy.LONGEST Pad to the longest sequence in the batch 268 | - PaddingStrategy.MAX_LENGTH: Pad to the max length (default) 269 | - PaddingStrategy.DO_NOT_PAD: Do not pad 270 | The tokenizer padding sides are defined in self.padding_side: 271 | 272 | - 'left': pads on the left of the sequences 273 | - 'right': pads on the right of the sequences 274 | pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value. 275 | This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability 276 | `>= 7.5` (Volta). 277 | return_attention_mask: 278 | (optional) Set to False to avoid returning attention mask (default: set to model specifics) 279 | """ 280 | # Load from model defaults 281 | assert self.padding_side == "left" 282 | 283 | required_input = encoded_inputs[self.model_input_names[0]] 284 | seq_length = len(required_input) 285 | 286 | if padding_strategy == PaddingStrategy.LONGEST: 287 | max_length = len(required_input) 288 | 289 | if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0): 290 | max_length = ((max_length // pad_to_multiple_of) + 291 | 1) * pad_to_multiple_of 292 | 293 | needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len( 294 | required_input) != max_length 295 | 296 | # Initialize attention mask if not present. 297 | if "attention_mask" not in encoded_inputs: 298 | encoded_inputs["attention_mask"] = [1] * seq_length 299 | 300 | if "position_ids" not in encoded_inputs: 301 | encoded_inputs["position_ids"] = list(range(seq_length)) 302 | 303 | if needs_to_be_padded: 304 | difference = max_length - len(required_input) 305 | 306 | if "attention_mask" in encoded_inputs: 307 | encoded_inputs["attention_mask"] = [0] * \ 308 | difference + encoded_inputs["attention_mask"] 309 | if "position_ids" in encoded_inputs: 310 | encoded_inputs["position_ids"] = [0] * \ 311 | difference + encoded_inputs["position_ids"] 312 | encoded_inputs[self.model_input_names[0]] = [ 313 | self.pad_token_id] * difference + required_input 314 | 315 | return encoded_inputs 316 | -------------------------------------------------------------------------------- /mz_kolors_core.py: -------------------------------------------------------------------------------- 1 | 2 | 3 | import gc 4 | import json 5 | import os 6 | import random 7 | import re 8 | import subprocess 9 | import sys 10 | from types import MethodType 11 | 12 | import torch 13 | import folder_paths 14 | import comfy.model_management as mm 15 | 16 | 17 | def chatglm3_text_encode(chatglm3_model, prompt): 18 | device = mm.get_torch_device() 19 | offload_device = mm.unet_offload_device() 20 | mm.unload_all_models() 21 | mm.soft_empty_cache() 22 | # Function to randomly select an option from the brackets 23 | 24 | def choose_random_option(match): 25 | options = match.group(1).split('|') 26 | return random.choice(options) 27 | 28 | prompt = re.sub(r'\{([^{}]*)\}', choose_random_option, prompt) 29 | 30 | # Define tokenizers and text encoders 31 | tokenizer = chatglm3_model['tokenizer'] 32 | text_encoder = chatglm3_model['text_encoder'] 33 | text_encoder.to(device) 34 | text_inputs = tokenizer( 35 | prompt, 36 | padding="max_length", 37 | max_length=256, 38 | truncation=True, 39 | return_tensors="pt", 40 | ).to(device) 41 | 42 | output = text_encoder( 43 | input_ids=text_inputs['input_ids'], 44 | attention_mask=text_inputs['attention_mask'], 45 | position_ids=text_inputs['position_ids'], 46 | output_hidden_states=True) 47 | 48 | # [batch_size, 77, 4096] 49 | prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone() 50 | text_proj = output.hidden_states[-1][-1, 51 | :, :].clone() # [batch_size, 4096] 52 | bs_embed, seq_len, _ = prompt_embeds.shape 53 | prompt_embeds = prompt_embeds.repeat(1, 1, 1) 54 | prompt_embeds = prompt_embeds.view( 55 | bs_embed, seq_len, -1) 56 | 57 | bs_embed = text_proj.shape[0] 58 | text_proj = text_proj.repeat(1, 1).view( 59 | bs_embed, -1 60 | ) 61 | text_encoder.to(offload_device) 62 | mm.soft_empty_cache() 63 | gc.collect() 64 | return prompt_embeds, text_proj 65 | 66 | 67 | def MZ_ChatGLM3Loader_call(args): 68 | # from .mz_kolors_utils import Utils 69 | # llm_dir = os.path.join(Utils.get_models_path(), "LLM") 70 | chatglm3_checkpoint = args.get("chatglm3_checkpoint") 71 | 72 | chatglm3_checkpoint_path = folder_paths.get_full_path( 73 | 'LLM', chatglm3_checkpoint) 74 | 75 | if not os.path.exists(chatglm3_checkpoint_path): 76 | raise RuntimeError( 77 | f"ERROR: Could not find chatglm3 checkpoint: {chatglm3_checkpoint_path}") 78 | 79 | from .chatglm3.configuration_chatglm import ChatGLMConfig 80 | from .chatglm3.modeling_chatglm import ChatGLMModel 81 | from .chatglm3.tokenization_chatglm import ChatGLMTokenizer 82 | 83 | offload_device = mm.unet_offload_device() 84 | 85 | text_encoder_config = os.path.join( 86 | os.path.dirname(__file__), 'configs', 'text_encoder_config.json') 87 | with open(text_encoder_config, 'r') as file: 88 | config = json.load(file) 89 | 90 | text_encoder_config = ChatGLMConfig(**config) 91 | 92 | from comfy.utils import load_torch_file 93 | from contextlib import nullcontext 94 | is_accelerate_available = False 95 | try: 96 | from accelerate import init_empty_weights 97 | from accelerate.utils import set_module_tensor_to_device 98 | is_accelerate_available = True 99 | except: 100 | pass 101 | 102 | with (init_empty_weights() if is_accelerate_available else nullcontext()): 103 | with torch.no_grad(): 104 | # 打印版本号 105 | print("torch version:", torch.__version__) 106 | text_encoder = ChatGLMModel(text_encoder_config).eval() 107 | if '4bit' in chatglm3_checkpoint: 108 | try: 109 | import cpm_kernels 110 | except ImportError: 111 | print("Installing cpm_kernels...") 112 | subprocess.run( 113 | [sys.executable, "-m", "pip", "install", "cpm_kernels"], check=True) 114 | pass 115 | text_encoder.quantize(4) 116 | elif '8bit' in chatglm3_checkpoint: 117 | text_encoder.quantize(8) 118 | text_encoder_sd = load_torch_file(chatglm3_checkpoint_path) 119 | if is_accelerate_available: 120 | for key in text_encoder_sd: 121 | set_module_tensor_to_device( 122 | text_encoder, key, device=offload_device, value=text_encoder_sd[key]) 123 | else: 124 | print("WARNING: Accelerate not available, use load_state_dict load model") 125 | text_encoder.load_state_dict(text_encoder_sd) 126 | 127 | tokenizer_path = os.path.join( 128 | os.path.dirname(__file__), 'configs', "tokenizer") 129 | tokenizer = ChatGLMTokenizer.from_pretrained(tokenizer_path) 130 | 131 | return ({"text_encoder": text_encoder, "tokenizer": tokenizer},) 132 | 133 | 134 | def MZ_ChatGLM3TextEncodeV2_call(args): 135 | text = args.get("text") 136 | chatglm3_model = args.get("chatglm3_model") 137 | prompt_embeds, pooled_output = chatglm3_text_encode( 138 | chatglm3_model, 139 | text, 140 | ) 141 | extra_kwargs = { 142 | "pooled_output": pooled_output, 143 | } 144 | extra_cond_keys = [ 145 | "width", 146 | "height", 147 | "crop_w", 148 | "crop_h", 149 | "target_width", 150 | "target_height" 151 | ] 152 | for key, value in args.items(): 153 | if key in extra_cond_keys: 154 | extra_kwargs[key] = value 155 | return ([[ 156 | prompt_embeds, 157 | # {"pooled_output": pooled_output}, 158 | extra_kwargs 159 | ]], ) 160 | 161 | 162 | def MZ_ChatGLM3Embeds2Conditioning_call(args): 163 | kolors_embeds = args.get("kolors_embeds") 164 | 165 | # kolors_embeds = { 166 | # 'prompt_embeds': prompt_embeds, 167 | # 'negative_prompt_embeds': negative_prompt_embeds, 168 | # 'pooled_prompt_embeds': text_proj, 169 | # 'negative_pooled_prompt_embeds': negative_text_proj 170 | # } 171 | 172 | positive = [[ 173 | kolors_embeds['prompt_embeds'], 174 | { 175 | "pooled_output": kolors_embeds['pooled_prompt_embeds'], 176 | "width": args.get("width"), 177 | "height": args.get("height"), 178 | "crop_w": args.get("crop_w"), 179 | "crop_h": args.get("crop_h"), 180 | "target_width": args.get("target_width"), 181 | "target_height": args.get("target_height") 182 | } 183 | ]] 184 | 185 | negative = [[ 186 | kolors_embeds['negative_prompt_embeds'], 187 | { 188 | "pooled_output": kolors_embeds['negative_pooled_prompt_embeds'], 189 | } 190 | ]] 191 | 192 | return (positive, negative) 193 | 194 | 195 | def MZ_KolorsUNETLoaderV2_call(kwargs): 196 | 197 | from . import hook_comfyui_kolors_v2 198 | import comfy.sd 199 | 200 | with hook_comfyui_kolors_v2.apply_kolors(): 201 | unet_name = kwargs.get("unet_name") 202 | unet_path = folder_paths.get_full_path("unet", unet_name) 203 | import comfy.utils 204 | sd = comfy.utils.load_torch_file(unet_path) 205 | model = comfy.sd.load_unet_state_dict(sd) 206 | if model is None: 207 | raise RuntimeError( 208 | "ERROR: Could not detect model type of: {}".format(unet_path)) 209 | 210 | return (model, ) 211 | 212 | 213 | def MZ_KolorsCheckpointLoaderSimple_call(kwargs): 214 | checkpoint_name = kwargs.get("ckpt_name") 215 | 216 | ckpt_path = folder_paths.get_full_path("checkpoints", checkpoint_name) 217 | 218 | from . import hook_comfyui_kolors_v2 219 | import comfy.sd 220 | 221 | with hook_comfyui_kolors_v2.apply_kolors(): 222 | out = comfy.sd.load_checkpoint_guess_config( 223 | ckpt_path, output_vae=True, output_clip=False, embedding_directory=folder_paths.get_folder_paths("embeddings")) 224 | 225 | unet, _, vae = out[:3] 226 | return (unet, vae) 227 | 228 | 229 | from comfy.cldm.cldm import ControlNet 230 | from comfy.controlnet import ControlLora 231 | 232 | 233 | def MZ_KolorsControlNetLoader_call(kwargs): 234 | control_net_name = kwargs.get("control_net_name") 235 | controlnet_path = folder_paths.get_full_path( 236 | "controlnet", control_net_name) 237 | 238 | from torch import nn 239 | from . import hook_comfyui_kolors_v2 240 | import comfy.controlnet 241 | 242 | with hook_comfyui_kolors_v2.apply_kolors(): 243 | control_net = comfy.controlnet.load_controlnet(controlnet_path) 244 | return (control_net, ) 245 | 246 | 247 | def MZ_KolorsControlNetPatch_call(kwargs): 248 | import copy 249 | from . import hook_comfyui_kolors_v2 250 | import comfy.model_management 251 | import comfy.model_patcher 252 | 253 | model = kwargs.get("model") 254 | control_net = kwargs.get("control_net") 255 | 256 | if hasattr(control_net, "control_model") and hasattr(control_net.control_model, "encoder_hid_proj"): 257 | return (control_net,) 258 | 259 | control_net = copy.deepcopy(control_net) 260 | 261 | import comfy.controlnet 262 | if isinstance(control_net, ControlLora): 263 | del_keys = [] 264 | for k in control_net.control_weights: 265 | if k.startswith("label_emb.0.0."): 266 | del_keys.append(k) 267 | 268 | for k in del_keys: 269 | control_net.control_weights.pop(k) 270 | 271 | super_pre_run = ControlLora.pre_run 272 | super_forward = ControlNet.forward 273 | 274 | def KolorsControlNet_forward(self, x, hint, timesteps, context, **kwargs): 275 | with torch.cuda.amp.autocast(enabled=True): 276 | context = self.encoder_hid_proj(context) 277 | return super_forward(self, x, hint, timesteps, context, **kwargs) 278 | 279 | def KolorsControlLora_pre_run(self, *args, **kwargs): 280 | result = super_pre_run(self, *args, **kwargs) 281 | 282 | if hasattr(self, "control_model"): 283 | if hasattr(self.control_model, "encoder_hid_proj"): 284 | return result 285 | 286 | setattr(self.control_model, "encoder_hid_proj", 287 | model.model.diffusion_model.encoder_hid_proj) 288 | 289 | self.control_model.forward = MethodType( 290 | KolorsControlNet_forward, self.control_model) 291 | 292 | return result 293 | 294 | control_net.pre_run = MethodType( 295 | KolorsControlLora_pre_run, control_net) 296 | 297 | super_copy = ControlLora.copy 298 | 299 | def KolorsControlLora_copy(self, *args, **kwargs): 300 | c = super_copy(self, *args, **kwargs) 301 | c.pre_run = MethodType( 302 | KolorsControlLora_pre_run, c) 303 | return c 304 | 305 | control_net.copy = MethodType( 306 | KolorsControlLora_copy, control_net) 307 | 308 | control_net = copy.deepcopy(control_net) 309 | 310 | elif isinstance(control_net, comfy.controlnet.ControlNet): 311 | model_label_emb = model.model.diffusion_model.label_emb 312 | 313 | control_net.control_model.label_emb = model_label_emb 314 | setattr(control_net.control_model, "encoder_hid_proj", 315 | model.model.diffusion_model.encoder_hid_proj) 316 | 317 | control_net.control_model_wrapped = comfy.model_patcher.ModelPatcher( 318 | control_net.control_model, load_device=control_net.load_device, offload_device=comfy.model_management.unet_offload_device()) 319 | 320 | super_forward = ControlNet.forward 321 | 322 | def KolorsControlNet_forward(self, x, hint, timesteps, context, **kwargs): 323 | with torch.cuda.amp.autocast(enabled=True): 324 | context = self.encoder_hid_proj(context) 325 | return super_forward(self, x, hint, timesteps, context, **kwargs) 326 | 327 | control_net.control_model.forward = MethodType( 328 | KolorsControlNet_forward, control_net.control_model) 329 | 330 | else: 331 | raise NotImplementedError( 332 | f"Type {control_net} not supported for KolorsControlNetPatch") 333 | 334 | return (control_net,) 335 | 336 | 337 | def MZ_KolorsCLIPVisionLoader_call(kwargs): 338 | import comfy.clip_vision 339 | from . import hook_comfyui_kolors_v2 340 | clip_name = kwargs.get("clip_name") 341 | clip_path = folder_paths.get_full_path("clip_vision", clip_name) 342 | with hook_comfyui_kolors_v2.apply_kolors(): 343 | clip_vision = comfy.clip_vision.load(clip_path) 344 | return (clip_vision,) 345 | 346 | 347 | def MZ_ApplySDXLSamplingSettings_call(kwargs): 348 | model = kwargs.get("model").clone() 349 | 350 | import comfy.model_sampling 351 | sampling_base = comfy.model_sampling.ModelSamplingDiscrete 352 | sampling_type = comfy.model_sampling.EPS 353 | 354 | class SDXLSampling(sampling_base, sampling_type): 355 | pass 356 | 357 | model.model.model_config.sampling_settings["beta_schedule"] = "linear" 358 | model.model.model_config.sampling_settings["linear_start"] = 0.00085 359 | model.model.model_config.sampling_settings["linear_end"] = 0.012 360 | model.model.model_config.sampling_settings["timesteps"] = 1000 361 | 362 | model_sampling = SDXLSampling(model.model.model_config) 363 | 364 | model.add_object_patch("model_sampling", model_sampling) 365 | 366 | return (model,) 367 | 368 | 369 | def MZ_ApplyCUDAGenerator_call(kwargs): 370 | model = kwargs.get("model") 371 | 372 | def prepare_noise(latent_image, seed, noise_inds=None): 373 | """ 374 | creates random noise given a latent image and a seed. 375 | optional arg skip can be used to skip and discard x number of noise generations for a given seed 376 | """ 377 | generator = torch.Generator(device="cuda").manual_seed(seed) 378 | if noise_inds is None: 379 | return torch.randn(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, generator=generator, device="cuda") 380 | 381 | unique_inds, inverse = np.unique(noise_inds, return_inverse=True) 382 | noises = [] 383 | for i in range(unique_inds[-1] + 1): 384 | noise = torch.randn([1] + list(latent_image.size())[1:], dtype=latent_image.dtype, 385 | layout=latent_image.layout, generator=generator, device="cuda") 386 | if i in unique_inds: 387 | noises.append(noise) 388 | noises = [noises[i] for i in inverse] 389 | noises = torch.cat(noises, axis=0) 390 | return noises 391 | 392 | import comfy.sample 393 | comfy.sample.prepare_noise = prepare_noise 394 | return (model,) 395 | -------------------------------------------------------------------------------- /chatglm3/quantization.py: -------------------------------------------------------------------------------- 1 | from torch.nn import Linear 2 | from torch.nn.parameter import Parameter 3 | 4 | import bz2 5 | import torch 6 | import base64 7 | import ctypes 8 | from transformers.utils import logging 9 | 10 | from typing import List 11 | from functools import partial 12 | 13 | logger = logging.get_logger(__name__) 14 | 15 | try: 16 | from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up 17 | 18 | import cpm_kernels.library.base 19 | 20 | original_windows_find_lib = cpm_kernels.library.base.windows_find_lib 21 | 22 | def windows_find_lib(name): 23 | result = original_windows_find_lib(name) 24 | if result is not None: 25 | return result 26 | import torch 27 | import os 28 | torch_dir = os.path.dirname(torch.__file__) 29 | torch_lib_dir = os.path.join(torch_dir, "lib") 30 | 31 | for name in os.listdir(torch_lib_dir): 32 | if name.startswith(name) and name.lower().endswith(".dll"): 33 | return os.path.join(torch_lib_dir, name) 34 | 35 | return None 36 | cpm_kernels.library.base.windows_find_lib = windows_find_lib 37 | 38 | class Kernel: 39 | def __init__(self, code: bytes, function_names: List[str]): 40 | self.code = code 41 | self._function_names = function_names 42 | self._cmodule = LazyKernelCModule(self.code) 43 | 44 | for name in self._function_names: 45 | setattr(self, name, KernelFunction(self._cmodule, name)) 46 | 47 | quantization_code = "$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" 48 | 49 | kernels = Kernel( 50 | bz2.decompress(base64.b64decode(quantization_code)), 51 | [ 52 | "int4WeightCompression", 53 | "int4WeightExtractionFloat", 54 | "int4WeightExtractionHalf", 55 | "int8WeightExtractionFloat", 56 | "int8WeightExtractionHalf", 57 | ], 58 | ) 59 | except Exception as exception: 60 | kernels = None 61 | logger.warning("Failed to load cpm_kernels:" + str(exception)) 62 | 63 | 64 | class W8A16Linear(torch.autograd.Function): 65 | @staticmethod 66 | def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width): 67 | ctx.inp_shape = inp.size() 68 | ctx.weight_bit_width = weight_bit_width 69 | out_features = quant_w.size(0) 70 | inp = inp.contiguous().view(-1, inp.size(-1)) 71 | weight = extract_weight_to_half(quant_w, scale_w, weight_bit_width) 72 | ctx.weight_shape = weight.size() 73 | output = inp.mm(weight.t()) 74 | ctx.save_for_backward(inp, quant_w, scale_w) 75 | return output.view(*(ctx.inp_shape[:-1] + (out_features,))) 76 | 77 | @staticmethod 78 | def backward(ctx, grad_output: torch.Tensor): 79 | inp, quant_w, scale_w = ctx.saved_tensors 80 | weight = extract_weight_to_half(quant_w, scale_w, ctx.weight_bit_width) 81 | grad_output = grad_output.contiguous().view(-1, weight.size(0)) 82 | grad_input = grad_output.mm(weight) 83 | grad_weight = grad_output.t().mm(inp) 84 | return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None, None 85 | 86 | 87 | def compress_int4_weight(weight: torch.Tensor): # (n, m) 88 | with torch.cuda.device(weight.device): 89 | n, m = weight.size(0), weight.size(1) 90 | assert m % 2 == 0 91 | m = m // 2 92 | out = torch.empty(n, m, dtype=torch.int8, device="cuda") 93 | stream = torch.cuda.current_stream() 94 | 95 | gridDim = (n, 1, 1) 96 | blockDim = (min(round_up(m, 32), 1024), 1, 1) 97 | 98 | kernels.int4WeightCompression( 99 | gridDim, 100 | blockDim, 101 | 0, 102 | stream, 103 | [ctypes.c_void_p(weight.data_ptr()), ctypes.c_void_p(out.data_ptr()), ctypes.c_int32(n), ctypes.c_int32(m)], 104 | ) 105 | return out 106 | 107 | 108 | def extract_weight_to_half(weight: torch.Tensor, scale_list: torch.Tensor, source_bit_width: int): 109 | assert scale_list.dtype in [torch.half, torch.bfloat16] 110 | assert weight.dtype in [torch.int8] 111 | if source_bit_width == 8: 112 | return weight.to(scale_list.dtype) * scale_list[:, None] 113 | elif source_bit_width == 4: 114 | func = ( 115 | kernels.int4WeightExtractionHalf if scale_list.dtype == torch.half else kernels.int4WeightExtractionBFloat16 116 | ) 117 | else: 118 | assert False, "Unsupported bit-width" 119 | 120 | with torch.cuda.device(weight.device): 121 | n, m = weight.size(0), weight.size(1) 122 | out = torch.empty(n, m * (8 // source_bit_width), dtype=scale_list.dtype, device="cuda") 123 | stream = torch.cuda.current_stream() 124 | 125 | gridDim = (n, 1, 1) 126 | blockDim = (min(round_up(m, 32), 1024), 1, 1) 127 | 128 | func( 129 | gridDim, 130 | blockDim, 131 | 0, 132 | stream, 133 | [ 134 | ctypes.c_void_p(weight.data_ptr()), 135 | ctypes.c_void_p(scale_list.data_ptr()), 136 | ctypes.c_void_p(out.data_ptr()), 137 | ctypes.c_int32(n), 138 | ctypes.c_int32(m), 139 | ], 140 | ) 141 | return out 142 | 143 | 144 | class QuantizedLinear(torch.nn.Module): 145 | def __init__(self, weight_bit_width: int, weight, bias=None, device="cpu", dtype=None, empty_init=False, *args, 146 | **kwargs): 147 | super().__init__() 148 | self.weight_bit_width = weight_bit_width 149 | 150 | shape = weight.shape 151 | 152 | if weight is None or empty_init: 153 | self.weight = torch.empty(shape[0], shape[1] * weight_bit_width // 8, dtype=torch.int8, device=device) 154 | self.weight_scale = torch.empty(shape[0], dtype=dtype, device=device) 155 | else: 156 | self.weight_scale = weight.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1) 157 | self.weight = torch.round(weight / self.weight_scale[:, None]).to(torch.int8) 158 | if weight_bit_width == 4: 159 | self.weight = compress_int4_weight(self.weight) 160 | 161 | try: 162 | self.weight = Parameter(self.weight.to(device), requires_grad=False) 163 | except: 164 | self.weight.to(device, dtype=self.weight.dtype) 165 | self.weight_scale = Parameter(self.weight_scale.to(device), requires_grad=False) 166 | self.bias = Parameter(bias.to(device), requires_grad=False) if bias is not None else None 167 | 168 | def forward(self, input): 169 | output = W8A16Linear.apply(input, self.weight, self.weight_scale, self.weight_bit_width) 170 | if self.bias is not None: 171 | output = output + self.bias 172 | return output 173 | 174 | 175 | def quantize(model, weight_bit_width, empty_init=False, device=None): 176 | """Replace fp16 linear with quantized linear""" 177 | for layer in model.layers: 178 | layer.self_attention.query_key_value = QuantizedLinear( 179 | weight_bit_width=weight_bit_width, 180 | weight=layer.self_attention.query_key_value.weight.to(torch.cuda.current_device()), 181 | bias=layer.self_attention.query_key_value.bias, 182 | dtype=layer.self_attention.query_key_value.weight.dtype, 183 | device=layer.self_attention.query_key_value.weight.device if device is None else device, 184 | empty_init=empty_init 185 | ) 186 | layer.self_attention.dense = QuantizedLinear( 187 | weight_bit_width=weight_bit_width, 188 | weight=layer.self_attention.dense.weight.to(torch.cuda.current_device()), 189 | bias=layer.self_attention.dense.bias, 190 | dtype=layer.self_attention.dense.weight.dtype, 191 | device=layer.self_attention.dense.weight.device if device is None else device, 192 | empty_init=empty_init 193 | ) 194 | layer.mlp.dense_h_to_4h = QuantizedLinear( 195 | weight_bit_width=weight_bit_width, 196 | weight=layer.mlp.dense_h_to_4h.weight.to(torch.cuda.current_device()), 197 | bias=layer.mlp.dense_h_to_4h.bias, 198 | dtype=layer.mlp.dense_h_to_4h.weight.dtype, 199 | device=layer.mlp.dense_h_to_4h.weight.device if device is None else device, 200 | empty_init=empty_init 201 | ) 202 | layer.mlp.dense_4h_to_h = QuantizedLinear( 203 | weight_bit_width=weight_bit_width, 204 | weight=layer.mlp.dense_4h_to_h.weight.to(torch.cuda.current_device()), 205 | bias=layer.mlp.dense_4h_to_h.bias, 206 | dtype=layer.mlp.dense_4h_to_h.weight.dtype, 207 | device=layer.mlp.dense_4h_to_h.weight.device if device is None else device, 208 | empty_init=empty_init 209 | ) 210 | 211 | return model 212 | -------------------------------------------------------------------------------- /hook_comfyui_kolors_v2.py: -------------------------------------------------------------------------------- 1 | import os 2 | from types import MethodType 3 | import warnings 4 | from comfy.model_detection import * 5 | import comfy.model_detection as model_detection 6 | import comfy.supported_models 7 | import comfy.utils 8 | 9 | import torch 10 | from comfy import model_base 11 | from comfy.model_base import sdxl_pooled, CLIPEmbeddingNoiseAugmentation, Timestep, ModelType 12 | 13 | 14 | from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel 15 | from comfy.cldm.cldm import ControlNet 16 | 17 | # try: 18 | # import comfy.samplers as samplers 19 | # original_CFGGuider_inner_set_conds = samplers.CFGGuider.set_conds 20 | 21 | # def patched_set_conds(self, positive, negative): 22 | # if isinstance(self.model_patcher.model, KolorsSDXL): 23 | # import copy 24 | # if "control" in positive[0][1]: 25 | # if hasattr(positive[0][1]["control"], "control_model"): 26 | # if positive[0][1]["control"].control_model.label_emb.shape[1] == 5632: 27 | # return 28 | 29 | 30 | # warnings.warn("该方法不再维护") 31 | # positive = copy.deepcopy(positive) 32 | # negative = copy.deepcopy(negative) 33 | # hid_proj = self.model_patcher.model.encoder_hid_proj 34 | # if hid_proj is not None: 35 | # positive[0][0] = hid_proj(positive[0][0]) 36 | # negative[0][0] = hid_proj(negative[0][0]) 37 | 38 | # if "control" in positive[0][1]: 39 | # if hasattr(positive[0][1]["control"], "control_model"): 40 | # positive[0][1]["control"].control_model.label_emb = self.model_patcher.model.diffusion_model.label_emb 41 | 42 | # if "control" in negative[0][1]: 43 | # if hasattr(negative[0][1]["control"], "control_model"): 44 | # negative[0][1]["control"].control_model.label_emb = self.model_patcher.model.diffusion_model.label_emb 45 | 46 | # return original_CFGGuider_inner_set_conds(self, positive, negative) 47 | 48 | # samplers.CFGGuider.set_conds = patched_set_conds 49 | # except ImportError: 50 | # print("CFGGuider not found, skipping patching") 51 | 52 | 53 | class KolorsUNetModel(UNetModel): 54 | def __init__(self, *args, **kwargs): 55 | super().__init__(*args, **kwargs) 56 | self.encoder_hid_proj = nn.Linear( 57 | 4096, 2048, bias=True) 58 | 59 | def forward(self, *args, **kwargs): 60 | with torch.cuda.amp.autocast(enabled=True): 61 | if "context" in kwargs: 62 | kwargs["context"] = self.encoder_hid_proj( 63 | kwargs["context"]) 64 | 65 | # if "y" in kwargs: 66 | # if kwargs["y"].shape[1] == 2816: 67 | # # 扩展至5632 68 | # kwargs["y"] = torch.cat( 69 | # torch.zeros(kwargs["y"].shape[0], 2816).to(kwargs["y"].device), kwargs["y"], dim=1) 70 | 71 | result = super().forward(*args, **kwargs) 72 | return result 73 | 74 | 75 | class KolorsSDXL(model_base.SDXL): 76 | def __init__(self, model_config, model_type=ModelType.EPS, device=None): 77 | model_config.sampling_settings["beta_schedule"] = "linear" 78 | model_config.sampling_settings["linear_start"] = 0.00085 79 | model_config.sampling_settings["linear_end"] = 0.014 80 | model_config.sampling_settings["timesteps"] = 1100 81 | model_type = ModelType.EPS 82 | model_base.BaseModel.__init__( 83 | self, model_config, model_type, device=device, unet_model=KolorsUNetModel) 84 | self.embedder = Timestep(256) 85 | self.noise_augmentor = CLIPEmbeddingNoiseAugmentation( 86 | **{"noise_schedule_config": {"timesteps": 1100, "beta_schedule": "linear", "linear_start": 0.00085, "linear_end": 0.014}, "timestep_dim": 1280}) 87 | 88 | def encode_adm(self, **kwargs): 89 | clip_pooled = sdxl_pooled(kwargs, self.noise_augmentor) 90 | width = kwargs.get("width", 768) 91 | height = kwargs.get("height", 768) 92 | crop_w = kwargs.get("crop_w", 0) 93 | crop_h = kwargs.get("crop_h", 0) 94 | target_width = kwargs.get("target_width", width) 95 | target_height = kwargs.get("target_height", height) 96 | 97 | out = [] 98 | out.append(self.embedder(torch.Tensor([height]))) 99 | out.append(self.embedder(torch.Tensor([width]))) 100 | out.append(self.embedder(torch.Tensor([crop_h]))) 101 | out.append(self.embedder(torch.Tensor([crop_w]))) 102 | out.append(self.embedder(torch.Tensor([target_height]))) 103 | out.append(self.embedder(torch.Tensor([target_width]))) 104 | flat = torch.flatten(torch.cat(out)).unsqueeze( 105 | dim=0).repeat(clip_pooled.shape[0], 1) 106 | return torch.cat((clip_pooled.to(flat.device), flat), dim=1) 107 | 108 | 109 | class KolorsSupported(comfy.supported_models.SDXL): 110 | unet_config = { 111 | "model_channels": 320, 112 | "use_linear_in_transformer": True, 113 | "transformer_depth": [0, 0, 2, 2, 10, 10], 114 | "context_dim": 2048, 115 | "adm_in_channels": 5632, 116 | "use_temporal_attention": False, 117 | } 118 | 119 | def get_model(self, state_dict, prefix="", device=None): 120 | out = KolorsSDXL(self, model_type=self.model_type( 121 | state_dict, prefix), device=device,) 122 | out.__class__ = model_base.SDXL 123 | if self.inpaint_model(): 124 | out.set_inpaint() 125 | return out 126 | 127 | 128 | def kolors_unet_config_from_diffusers_unet(state_dict, dtype=None): 129 | match = {} 130 | transformer_depth = [] 131 | 132 | attn_res = 1 133 | down_blocks = count_blocks(state_dict, "down_blocks.{}") 134 | for i in range(down_blocks): 135 | attn_blocks = count_blocks( 136 | state_dict, "down_blocks.{}.attentions.".format(i) + '{}') 137 | res_blocks = count_blocks( 138 | state_dict, "down_blocks.{}.resnets.".format(i) + '{}') 139 | for ab in range(attn_blocks): 140 | transformer_count = count_blocks( 141 | state_dict, "down_blocks.{}.attentions.{}.transformer_blocks.".format(i, ab) + '{}') 142 | transformer_depth.append(transformer_count) 143 | if transformer_count > 0: 144 | match["context_dim"] = state_dict["down_blocks.{}.attentions.{}.transformer_blocks.0.attn2.to_k.weight".format( 145 | i, ab)].shape[1] 146 | 147 | attn_res *= 2 148 | if attn_blocks == 0: 149 | for i in range(res_blocks): 150 | transformer_depth.append(0) 151 | 152 | match["transformer_depth"] = transformer_depth 153 | 154 | match["model_channels"] = state_dict["conv_in.weight"].shape[0] 155 | match["in_channels"] = state_dict["conv_in.weight"].shape[1] 156 | match["adm_in_channels"] = None 157 | if "class_embedding.linear_1.weight" in state_dict: 158 | match["adm_in_channels"] = state_dict["class_embedding.linear_1.weight"].shape[1] 159 | elif "add_embedding.linear_1.weight" in state_dict: 160 | match["adm_in_channels"] = state_dict["add_embedding.linear_1.weight"].shape[1] 161 | 162 | Kolors = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 163 | 'num_classes': 'sequential', 'adm_in_channels': 5632, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 164 | 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10, 165 | 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], 166 | 'use_temporal_attention': False, 'use_temporal_resblock': False} 167 | 168 | Kolors_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 169 | 'num_classes': 'sequential', 'adm_in_channels': 5632, 'dtype': dtype, 'in_channels': 9, 'model_channels': 320, 170 | 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10, 171 | 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], 172 | 'use_temporal_attention': False, 'use_temporal_resblock': False} 173 | 174 | Kolors_ip2p = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 175 | 'num_classes': 'sequential', 'adm_in_channels': 5632, 'dtype': dtype, 'in_channels': 8, 'model_channels': 320, 176 | 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10, 177 | 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], 178 | 'use_temporal_attention': False, 'use_temporal_resblock': False} 179 | 180 | SDXL = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 181 | 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 182 | 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10, 183 | 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10], 184 | 'use_temporal_attention': False, 'use_temporal_resblock': False} 185 | 186 | SDXL_mid_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 187 | 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 188 | 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 0, 0, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 1, 189 | 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 0, 0, 0, 1, 1, 1], 190 | 'use_temporal_attention': False, 'use_temporal_resblock': False} 191 | 192 | SDXL_small_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 193 | 'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 194 | 'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 0, 0, 0, 0], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 0, 195 | 'use_linear_in_transformer': True, 'num_head_channels': 64, 'context_dim': 1, 'transformer_depth_output': [0, 0, 0, 0, 0, 0, 0, 0, 0], 196 | 'use_temporal_attention': False, 'use_temporal_resblock': False} 197 | 198 | supported_models = [Kolors, Kolors_inpaint, 199 | Kolors_ip2p, SDXL, SDXL_mid_cnet, SDXL_small_cnet] 200 | 201 | for unet_config in supported_models: 202 | matches = True 203 | for k in match: 204 | if match[k] != unet_config[k]: 205 | print("key {} does not match".format( 206 | k), match[k], "||", unet_config[k]) 207 | matches = False 208 | break 209 | if matches: 210 | return convert_config(unet_config) 211 | return None 212 | 213 | 214 | import comfy.ldm.modules.diffusionmodules.openaimodel 215 | from torch import nn 216 | 217 | 218 | def load_clipvision_336_from_sd(sd, prefix="", convert_keys=False): 219 | from comfy.clip_vision import ClipVisionModel, convert_to_transformers 220 | 221 | json_config = os.path.join(os.path.dirname( 222 | os.path.realpath(__file__)), "clip_vit_336", "config.json") 223 | 224 | clip = ClipVisionModel(json_config) 225 | 226 | m, u = clip.load_sd(sd) 227 | if len(m) > 0: 228 | logging.warning("missing clip vision: {}".format(m)) 229 | u = set(u) 230 | keys = list(sd.keys()) 231 | for k in keys: 232 | if k not in u: 233 | t = sd.pop(k) 234 | del t 235 | 236 | # def vis_forward(self, pixel_values, attention_mask=None, intermediate_output=None): 237 | # pixel_values = nn.functional.interpolate( 238 | # pixel_values, size=(336, 336), mode='bilinear', align_corners=False) 239 | # x = self.embeddings(pixel_values) 240 | # x = self.pre_layrnorm(x) 241 | # # TODO: attention_mask? 242 | # x, i = self.encoder( 243 | # x, mask=None, intermediate_output=intermediate_output) 244 | # pooled_output = self.post_layernorm(x[:, 0, :]) 245 | # return x, i, pooled_output 246 | 247 | # clip.model.vision_model.forward = MethodType( 248 | # vis_forward, clip.model.vision_model 249 | # ) 250 | 251 | return clip 252 | 253 | 254 | class KolorsControlNet(ControlNet): 255 | def __init__(self, *args, **kwargs): 256 | adm_in_channels = kwargs["adm_in_channels"] 257 | if adm_in_channels == 2816: 258 | # 异常: 该加载器不支持SDXL类型, 请使用ControlNet加载器+KolorsControlNetPatch节点 259 | raise Exception( 260 | "This loader does not support SDXL type, please use ControlNet loader + KolorsControlNetPatch node") 261 | 262 | super().__init__(*args, **kwargs) 263 | self.encoder_hid_proj = nn.Linear( 264 | 4096, 2048, bias=True) 265 | 266 | def forward(self, *args, **kwargs): 267 | with torch.cuda.amp.autocast(enabled=True): 268 | if "context" in kwargs: 269 | kwargs["context"] = self.encoder_hid_proj( 270 | kwargs["context"]) 271 | 272 | result = super().forward(*args, **kwargs) 273 | return result 274 | 275 | 276 | class apply_kolors: 277 | def __enter__(self): 278 | import comfy.ldm.modules.diffusionmodules.openaimodel 279 | import comfy.cldm.cldm 280 | import comfy.utils 281 | import comfy.clip_vision 282 | 283 | self.original_load_clipvision_from_sd = comfy.clip_vision.load_clipvision_from_sd 284 | comfy.clip_vision.load_clipvision_from_sd = load_clipvision_336_from_sd 285 | 286 | self.original_UNET_MAP_BASIC = comfy.utils.UNET_MAP_BASIC.copy() 287 | comfy.utils.UNET_MAP_BASIC.add( 288 | ("encoder_hid_proj.weight", "encoder_hid_proj.weight"), 289 | ) 290 | comfy.utils.UNET_MAP_BASIC.add( 291 | ("encoder_hid_proj.bias", "encoder_hid_proj.bias"), 292 | ) 293 | 294 | self.original_unet_config_from_diffusers_unet = model_detection.unet_config_from_diffusers_unet 295 | model_detection.unet_config_from_diffusers_unet = kolors_unet_config_from_diffusers_unet 296 | 297 | import comfy.supported_models 298 | self.original_supported_models = comfy.supported_models.models 299 | comfy.supported_models.models = [KolorsSupported] 300 | 301 | self.original_controlnet = comfy.cldm.cldm.ControlNet 302 | comfy.cldm.cldm.ControlNet = KolorsControlNet 303 | 304 | def __exit__(self, type, value, traceback): 305 | import comfy.ldm.modules.diffusionmodules.openaimodel 306 | import comfy.cldm.cldm 307 | import comfy.utils 308 | comfy.utils.UNET_MAP_BASIC = self.original_UNET_MAP_BASIC 309 | 310 | model_detection.unet_config_from_diffusers_unet = self.original_unet_config_from_diffusers_unet 311 | 312 | import comfy.supported_models 313 | comfy.supported_models.models = self.original_supported_models 314 | 315 | import comfy.clip_vision 316 | comfy.clip_vision.load_clipvision_from_sd = self.original_load_clipvision_from_sd 317 | 318 | comfy.cldm.cldm.ControlNet = self.original_controlnet 319 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /mz_kolors_utils.py: -------------------------------------------------------------------------------- 1 | 2 | import json 3 | import os 4 | import shutil 5 | import subprocess 6 | import sys 7 | import threading 8 | import time 9 | import numpy as np 10 | import folder_paths 11 | import base64 12 | from PIL import Image, ImageFilter 13 | import io 14 | import torch 15 | import re 16 | import hashlib 17 | import cv2 18 | # sys.path.append(os.path.join(os.path.dirname(__file__))) 19 | temp_directory = folder_paths.get_temp_directory() 20 | from tqdm import tqdm 21 | import requests 22 | import comfy.utils 23 | 24 | 25 | CACHE_POOL = {} 26 | 27 | 28 | class Utils: 29 | def Md5(str): 30 | return hashlib.md5(str.encode('utf-8')).hexdigest() 31 | 32 | def check_frames_path(frames_path): 33 | 34 | if frames_path == "" or frames_path.startswith(".") or frames_path.startswith("/") or frames_path.endswith("/") or frames_path.endswith("\\"): 35 | return "frames_path不能为空" 36 | 37 | frames_path = os.path.join( 38 | folder_paths.get_output_directory(), frames_path) 39 | 40 | if frames_path == folder_paths.get_output_directory(): 41 | return "frames_path不能为output目录" 42 | 43 | return "" 44 | 45 | def base64_to_pil_image(base64_str): 46 | if base64_str is None: 47 | return None 48 | if len(base64_str) == 0: 49 | return None 50 | if type(base64_str) not in [str, bytes]: 51 | return None 52 | if base64_str.startswith("data:image/png;base64,"): 53 | base64_str = base64_str.split(",")[-1] 54 | base64_str = base64_str.encode("utf-8") 55 | base64_str = base64.b64decode(base64_str) 56 | return Image.open(io.BytesIO(base64_str)) 57 | 58 | def pil_image_to_base64(pil_image): 59 | buffered = io.BytesIO() 60 | pil_image.save(buffered, format="PNG") 61 | img_str = base64.b64encode(buffered.getvalue()) 62 | img_str = str(img_str, encoding="utf-8") 63 | return f"data:image/png;base64,{img_str}" 64 | 65 | def listdir_png(path): 66 | try: 67 | files = os.listdir(path) 68 | new_files = [] 69 | for file in files: 70 | if file.endswith(".png"): 71 | new_files.append(file) 72 | files = new_files 73 | files.sort(key=lambda x: int(os.path.basename(x).split(".")[0])) 74 | return files 75 | except Exception as e: 76 | return [] 77 | 78 | def listdir_models(path): 79 | try: 80 | relative_paths = [] 81 | for root, dirs, files in os.walk(path): 82 | for file in files: 83 | relative_paths.append(os.path.relpath( 84 | os.path.join(root, file), path)) 85 | relative_paths = [f for f in relative_paths if f.endswith(".safetensors") or f.endswith( 86 | ".pt") or f.endswith(".pth") or f.endswith(".onnx")] 87 | return relative_paths 88 | 89 | except Exception as e: 90 | 91 | return [] 92 | 93 | def tensor2pil(image): 94 | return Image.fromarray(np.clip(255.0 * image.cpu().numpy().squeeze(), 0, 255).astype(np.uint8)) 95 | 96 | # Convert PIL to Tensor 97 | 98 | def pil2tensor(image): 99 | return torch.from_numpy(np.array(image).astype(np.float32) / 255.0).unsqueeze(0)[0] 100 | 101 | def pil2cv(image): 102 | return cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR) 103 | 104 | def cv2pil(image): 105 | return Image.fromarray(cv2.cvtColor(image, cv2.COLOR_BGR2RGB)) 106 | 107 | def list_tensor2tensor(data): 108 | result_tensor = torch.stack(data) 109 | return result_tensor 110 | 111 | def loadImage(path): 112 | img = Image.open(path) 113 | img = img.convert("RGB") 114 | return img 115 | 116 | def vae_encode_crop_pixels(pixels): 117 | x = (pixels.shape[1] // 8) * 8 118 | y = (pixels.shape[2] // 8) * 8 119 | if pixels.shape[1] != x or pixels.shape[2] != y: 120 | x_offset = (pixels.shape[1] % 8) // 2 121 | y_offset = (pixels.shape[2] % 8) // 2 122 | pixels = pixels[:, x_offset:x + x_offset, y_offset:y + y_offset, :] 123 | return pixels 124 | 125 | def native_vae_encode(vae, image): 126 | pixels = Utils.vae_encode_crop_pixels(image) 127 | t = vae.encode(pixels[:, :, :, :3]) 128 | return {"samples": t} 129 | 130 | def native_vae_encode_for_inpaint(vae, pixels, mask): 131 | x = (pixels.shape[1] // 8) * 8 132 | y = (pixels.shape[2] // 8) * 8 133 | mask = torch.nn.functional.interpolate(mask.reshape( 134 | (-1, 1, mask.shape[-2], mask.shape[-1])), size=(pixels.shape[1], pixels.shape[2]), mode="bilinear") 135 | 136 | pixels = pixels.clone() 137 | if pixels.shape[1] != x or pixels.shape[2] != y: 138 | x_offset = (pixels.shape[1] % 8) // 2 139 | y_offset = (pixels.shape[2] % 8) // 2 140 | pixels = pixels[:, x_offset:x + x_offset, y_offset:y + y_offset, :] 141 | mask = mask[:, :, x_offset:x + x_offset, y_offset:y + y_offset] 142 | 143 | # grow mask by a few pixels to keep things seamless in latent space 144 | 145 | mask_erosion = mask 146 | 147 | m = (1.0 - mask.round()).squeeze(1) 148 | for i in range(3): 149 | pixels[:, :, :, i] -= 0.5 150 | pixels[:, :, :, i] *= m 151 | pixels[:, :, :, i] += 0.5 152 | t = vae.encode(pixels) 153 | 154 | return {"samples": t, "noise_mask": (mask_erosion[:, :, :x, :y].round())} 155 | 156 | def native_vae_decode(vae, samples): 157 | return vae.decode(samples["samples"]) 158 | 159 | def native_clip_text_encode(clip, text): 160 | tokens = clip.tokenize(text) 161 | cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True) 162 | return [[cond, {"pooled_output": pooled}]] 163 | 164 | def a1111_clip_text_encode(clip, text): 165 | try: 166 | from . import ADV_CLIP_emb_encode 167 | cond, pooled = ADV_CLIP_emb_encode.advanced_encode( 168 | clip, text, "none", "A1111", w_max=1.0, apply_to_pooled=False) 169 | return [[cond, {"pooled_output": pooled}]] 170 | except Exception as e: 171 | import nodes 172 | return nodes.CLIPTextEncode().encode(clip, text)[0] 173 | 174 | def cache_get(key): 175 | return CACHE_POOL.get(key, None) 176 | 177 | def cache_set(key, value): 178 | global CACHE_POOL 179 | CACHE_POOL[key] = value 180 | return True 181 | 182 | def get_models_path(): 183 | return folder_paths.models_dir 184 | 185 | def get_gguf_models_path(): 186 | models_path = os.path.join( 187 | folder_paths.models_dir, "gguf") 188 | os.makedirs(models_path, exist_ok=True) 189 | return models_path 190 | 191 | def get_translate_object(from_code, to_code): 192 | try: 193 | is_disabel_argostranslate = Utils.cache_get( 194 | "is_disabel_argostranslate") 195 | 196 | if is_disabel_argostranslate is not None: 197 | return None 198 | 199 | try: 200 | import argostranslate 201 | from argostranslate import translate, package 202 | except ImportError: 203 | subprocess.run([ 204 | sys.executable, "-m", 205 | "pip", "install", "argostranslate"], check=True) 206 | 207 | try: 208 | import argostranslate 209 | from argostranslate import translate, package 210 | except ImportError: 211 | Utils.cache_set("is_disabel_argostranslate", True) 212 | print( 213 | "argostranslate not found and install failed , will disable it") 214 | return None 215 | 216 | packages = package.get_installed_packages() 217 | installed_packages = {} 218 | for p in packages: 219 | installed_packages[f"{p.from_code}_{p.to_code}"] = p 220 | 221 | argosmodel_dir = os.path.join( 222 | Utils.get_models_path(), "argosmodel") 223 | if not os.path.exists(argosmodel_dir): 224 | os.makedirs(argosmodel_dir) 225 | 226 | model_name = None 227 | if from_code == "zh" and to_code == "en": 228 | model_name = "zh_en" 229 | elif from_code == "en" and to_code == "zh": 230 | model_name = "en_zh" 231 | else: 232 | return None 233 | 234 | if Utils.cache_get(f"argostranslate_{model_name}") is not None: 235 | return Utils.cache_get(f"argostranslate_{model_name}") 236 | 237 | if installed_packages.get(model_name, None) is None: 238 | if not os.path.exists(os.path.join(argosmodel_dir, f"translate-{model_name}-1_9.argosmodel")): 239 | argosmodel_file = Utils.download_file( 240 | url=f"https://www.modelscope.cn/api/v1/models/wailovet/MinusZoneAIModels/repo?Revision=master&FilePath=argosmodel%2Ftranslate-{model_name}-1_9.argosmodel", 241 | filepath=os.path.join( 242 | argosmodel_dir, f"translate-{model_name}-1_9.argosmodel"), 243 | ) 244 | else: 245 | argosmodel_file = os.path.join( 246 | argosmodel_dir, f"translate-{model_name}-1_9.argosmodel") 247 | package.install_from_path(argosmodel_file) 248 | 249 | translate_object = translate.get_translation_from_codes( 250 | from_code=from_code, to_code=to_code) 251 | 252 | Utils.cache_set(f"argostranslate_{model_name}", translate_object) 253 | 254 | return translate_object 255 | except Exception as e: 256 | Utils.cache_set("is_disabel_argostranslate", True) 257 | print( 258 | "argostranslate not found and install failed , will disable it") 259 | print(f"get_translate_object error: {e}") 260 | return None 261 | 262 | def translate_text(text, from_code, to_code): 263 | translation = Utils.get_translate_object(from_code, to_code) 264 | if translation is None: 265 | return text 266 | 267 | # Translate 268 | translatedText = translation.translate( 269 | text) 270 | 271 | return translatedText 272 | 273 | def zh2en(text): 274 | try: 275 | return Utils.translate_text(text, "zh", "en") 276 | except Exception as e: 277 | print(f"zh2en error: {e}") 278 | return text 279 | 280 | def en2zh(text): 281 | try: 282 | return Utils.translate_text(text, "en", "zh") 283 | except Exception as e: 284 | print(f"en2zh error: {e}") 285 | return text 286 | 287 | def prompt_zh_to_en(prompt): 288 | prompt = prompt.replace("，", ",") 289 | prompt = prompt.replace("。", ",") 290 | prompt = prompt.replace("\n", ",") 291 | tags = prompt.split(",") 292 | # 判断是否有中文 293 | for i, tag in enumerate(tags): 294 | if re.search(u'[\u4e00-\u9fff]', tag): 295 | tags[i] = Utils.zh2en(tag) 296 | # 如果第一个字母是大写,转为小写 297 | if tags[i][0].isupper(): 298 | tags[i] = tags[i].lower().replace(".", "") 299 | 300 | return ",".join(tags) 301 | 302 | def mask_resize(mask, width, height): 303 | mask = mask.unsqueeze(0).unsqueeze(0) 304 | mask = torch.nn.functional.interpolate( 305 | mask, size=(height, width), mode="bilinear") 306 | mask = mask.squeeze(0).squeeze(0) 307 | return mask 308 | 309 | def mask_threshold(interested_mask): 310 | mask_image = Utils.tensor2pil(interested_mask) 311 | mask_image_cv2 = Utils.pil2cv(mask_image) 312 | ret, thresh1 = cv2.threshold( 313 | mask_image_cv2, 127, 255, cv2.THRESH_BINARY) 314 | thresh1 = Utils.cv2pil(thresh1) 315 | thresh1 = np.array(thresh1) 316 | thresh1 = thresh1[:, :, 0] 317 | return Utils.pil2tensor(thresh1) 318 | 319 | def mask_erode(interested_mask, value): 320 | value = int(value) 321 | mask_image = Utils.tensor2pil(interested_mask) 322 | mask_image_cv2 = Utils.pil2cv(mask_image) 323 | kernel = np.ones((5, 5), np.uint8) 324 | erosion = cv2.erode(mask_image_cv2, kernel, iterations=value) 325 | erosion = Utils.cv2pil(erosion) 326 | erosion = np.array(erosion) 327 | erosion = erosion[:, :, 0] 328 | return Utils.pil2tensor(erosion) 329 | 330 | def mask_dilate(interested_mask, value): 331 | value = int(value) 332 | mask_image = Utils.tensor2pil(interested_mask) 333 | mask_image_cv2 = Utils.pil2cv(mask_image) 334 | kernel = np.ones((5, 5), np.uint8) 335 | dilation = cv2.dilate(mask_image_cv2, kernel, iterations=value) 336 | dilation = Utils.cv2pil(dilation) 337 | dilation = np.array(dilation) 338 | dilation = dilation[:, :, 0] 339 | return Utils.pil2tensor(dilation) 340 | 341 | def mask_edge_opt(interested_mask, edge_feathering): 342 | 343 | mask_image = Utils.tensor2pil(interested_mask) 344 | mask_image_cv2 = Utils.pil2cv(mask_image) 345 | 346 | # 高斯模糊 347 | dilation2 = Utils.cv2pil(mask_image_cv2) 348 | dilation2 = mask_image.filter( 349 | ImageFilter.GaussianBlur(edge_feathering)) 350 | 351 | # mask_image dilation2 图片蒙版叠加 352 | dilation2 = Utils.pil2cv(dilation2) 353 | # dilation2[mask_image_cv2 < 127] = 0 354 | dilation2 = Utils.cv2pil(dilation2) 355 | # to RGB 356 | dilation2 = np.array(dilation2) 357 | dilation2 = dilation2[:, :, 0] 358 | return Utils.pil2tensor(dilation2) 359 | 360 | def mask_composite(destination, source, x, y, mask=None, multiplier=8, resize_source=False): 361 | source = source.to(destination.device) 362 | if resize_source: 363 | source = torch.nn.functional.interpolate(source, size=( 364 | destination.shape[2], destination.shape[3]), mode="bilinear") 365 | 366 | source = comfy.utils.repeat_to_batch_size(source, destination.shape[0]) 367 | 368 | x = max(-source.shape[3] * multiplier, 369 | min(x, destination.shape[3] * multiplier)) 370 | y = max(-source.shape[2] * multiplier, 371 | min(y, destination.shape[2] * multiplier)) 372 | 373 | left, top = (x // multiplier, y // multiplier) 374 | right, bottom = (left + source.shape[3], top + source.shape[2],) 375 | 376 | if mask is None: 377 | mask = torch.ones_like(source) 378 | else: 379 | mask = mask.to(destination.device, copy=True) 380 | mask = torch.nn.functional.interpolate(mask.reshape( 381 | (-1, 1, mask.shape[-2], mask.shape[-1])), size=(source.shape[2], source.shape[3]), mode="bilinear") 382 | mask = comfy.utils.repeat_to_batch_size(mask, source.shape[0]) 383 | 384 | # calculate the bounds of the source that will be overlapping the destination 385 | # this prevents the source trying to overwrite latent pixels that are out of bounds 386 | # of the destination 387 | visible_width, visible_height = ( 388 | destination.shape[3] - left + min(0, x), destination.shape[2] - top + min(0, y),) 389 | 390 | mask = mask[:, :, :visible_height, :visible_width] 391 | inverse_mask = torch.ones_like(mask) - mask 392 | 393 | source_portion = mask * source[:, :, :visible_height, :visible_width] 394 | destination_portion = inverse_mask * \ 395 | destination[:, :, top:bottom, left:right] 396 | 397 | destination[:, :, top:bottom, 398 | left:right] = source_portion + destination_portion 399 | return destination 400 | 401 | def latent_upscale_by(samples, scale_by): 402 | s = samples.copy() 403 | width = round(samples["samples"].shape[3] * scale_by) 404 | height = round(samples["samples"].shape[2] * scale_by) 405 | s["samples"] = comfy.utils.common_upscale( 406 | samples["samples"], width, height, "nearest-exact", "disabled") 407 | return s 408 | 409 | def resize_by(image, percent): 410 | # 判断类型是否为PIL 411 | if not isinstance(image, Image.Image): 412 | image = Image.fromarray(image) 413 | 414 | width, height = image.size 415 | new_width = int(width * percent) 416 | new_height = int(height * percent) 417 | return image.resize((new_width, new_height), Image.LANCZOS) 418 | 419 | def resize_max(im, dst_w, dst_h): 420 | src_w, src_h = im.size 421 | 422 | if src_h < src_w: 423 | newWidth = dst_w 424 | newHeight = dst_w * src_h // src_w 425 | else: 426 | newWidth = dst_h * src_w // src_h 427 | newHeight = dst_h 428 | 429 | newHeight = newHeight // 8 * 8 430 | newWidth = newWidth // 8 * 8 431 | 432 | return im.resize((newWidth, newHeight), Image.Resampling.LANCZOS) 433 | 434 | def get_device(): 435 | return comfy.model_management.get_torch_device() 436 | 437 | def download_small_file(url, filepath): 438 | response = requests.get(url) 439 | os.makedirs(os.path.dirname(filepath), exist_ok=True) 440 | with open(filepath, "wb") as f: 441 | f.write(response.content) 442 | return filepath 443 | 444 | def download_file(url, filepath, threads=8, retries=6): 445 | 446 | get_size_tmp = requests.get(url, stream=True) 447 | total_size = int(get_size_tmp.headers.get("content-length", 0)) 448 | 449 | print(f"Downloading {url} to {filepath} with size {total_size} bytes") 450 | 451 | # 如果文件大小小于 50MB,使用download_small_file 452 | if total_size < 50 * 1024 * 1024: 453 | return Utils.download_small_file(url, filepath) 454 | 455 | base_filename = os.path.basename(filepath) 456 | cache_dir = os.path.join(os.path.dirname( 457 | filepath), f"{base_filename}.t_{threads}_cache") 458 | os.makedirs(cache_dir, exist_ok=True) 459 | 460 | def get_total_existing_size(): 461 | fs = os.listdir(cache_dir) 462 | existing_size = 0 463 | for f in fs: 464 | if f.startswith("block_"): 465 | existing_size += os.path.getsize( 466 | os.path.join(cache_dir, f)) 467 | return existing_size 468 | 469 | total_existing_size = get_total_existing_size() 470 | 471 | if total_size != 0 and total_existing_size != total_size: 472 | 473 | with tqdm(total=total_size, initial=total_existing_size, unit="B", unit_scale=True) as progress_bar: 474 | all_threads = [] 475 | 476 | for i in range(threads): 477 | cache_filepath = os.path.join(cache_dir, f"block_{i}") 478 | 479 | start = total_size // threads * i 480 | end = total_size // threads * (i + 1) - 1 481 | 482 | if i == threads - 1: 483 | end = total_size 484 | 485 | # Check if the file already exists 486 | if os.path.exists(cache_filepath): 487 | # Get the size of the existing file 488 | existing_size = os.path.getsize(cache_filepath) 489 | else: 490 | existing_size = 0 491 | 492 | headers = {"Range": f"bytes={start + existing_size}-{end}"} 493 | if end == total_size: 494 | headers = {"Range": f"bytes={start + existing_size}-"} 495 | if start + existing_size >= end: 496 | continue 497 | # print(f"Downloading {cache_filepath} with headers bytes={start + existing_size}-{end}") 498 | 499 | # Streaming, so we can iterate over the response. 500 | response = requests.get(url, stream=True, headers=headers) 501 | 502 | def download_file_thread(response, cache_filepath): 503 | block_size = 1024 504 | if end - (start + existing_size) < block_size: 505 | block_size = end - (start + existing_size) 506 | with open(cache_filepath, "ab") as file: 507 | for data in response.iter_content(block_size): 508 | file.write(data) 509 | progress_bar.update( 510 | len(data) 511 | ) 512 | 513 | t = threading.Thread( 514 | target=download_file_thread, args=(response, cache_filepath)) 515 | 516 | all_threads.append(t) 517 | 518 | t.start() 519 | 520 | for t in all_threads: 521 | t.join() 522 | 523 | if total_size != 0 and get_total_existing_size() > total_size: 524 | # 文件下载失败 525 | shutil.rmtree(cache_dir) 526 | raise RuntimeError("Download failed, file is incomplete") 527 | 528 | if total_size != 0 and total_size != get_total_existing_size(): 529 | if retries > 0: 530 | retries -= 1 531 | print( 532 | f"Download failed: {total_size} != {get_total_existing_size()}, retrying... {retries} retries left") 533 | return Utils.download_file(url, filepath, threads, retries) 534 | 535 | # 文件损坏 536 | raise RuntimeError( 537 | f"Download failed: {total_size} != {get_total_existing_size()}") 538 | 539 | if os.path.exists(filepath): 540 | shutil.move(filepath, filepath + ".old." + 541 | time.strftime("%Y%m%d%H%M%S")) 542 | 543 | # merge the files 544 | with open(filepath, "wb") as f: 545 | for i in range(threads): 546 | cache_filepath = os.path.join(cache_dir, f"block_{i}") 547 | with open(cache_filepath, "rb") as cf: 548 | f.write(cf.read()) 549 | 550 | shutil.rmtree(cache_dir) 551 | return filepath 552 | 553 | def hf_download_model(url, only_get_path=False): 554 | if not url.startswith("https://"): 555 | raise ValueError("URL must start with https://") 556 | if url.startswith("https://huggingface.co/") or url.startswith("https://hf-mirror.com/"): 557 | base_model_path = os.path.abspath(os.path.join( 558 | Utils.get_models_path(), "transformers_models")) 559 | # https://huggingface.co/FaradayDotDev/llama-3-8b-Instruct-GGUF/resolve/main/llama-3-8b-Instruct.Q2_K.gguf?download=true 560 | texts = url.split("?")[0].split("/") 561 | file_name = texts[-1] 562 | zone_path = f"{texts[3]}/{texts[4]}" 563 | 564 | save_path = os.path.join(base_model_path, zone_path, file_name) 565 | 566 | if os.path.exists(save_path) is False: 567 | if only_get_path: 568 | return None 569 | os.makedirs(os.path.join( 570 | base_model_path, zone_path), exist_ok=True) 571 | Utils.download_file(url, save_path) 572 | 573 | # Utils.print_log( 574 | # f"File {save_path} => {os.path.getsize(save_path)} ") 575 | 576 | # 获取大小 577 | if os.path.getsize(save_path) == 0: 578 | if only_get_path: 579 | return None 580 | os.remove(save_path) 581 | raise ValueError(f"Download failed: {url}") 582 | return save_path 583 | else: 584 | texts = url.split("?")[0].split("/") 585 | host = texts[2].replace(".", "_") 586 | base_model_path = os.path.abspath(os.path.join( 587 | Utils.get_models_path(), f"{host}_models")) 588 | 589 | file_name = texts[-1] 590 | file_name_no_ext = os.path.splitext(file_name)[0] 591 | file_ext = os.path.splitext(file_name)[1] 592 | md5_hash = Utils.Md5(url) 593 | 594 | save_path = os.path.join( 595 | base_model_path, f"{file_name_no_ext}.{md5_hash}{file_ext}") 596 | 597 | if os.path.exists(save_path) is False: 598 | if only_get_path: 599 | return None 600 | os.makedirs(base_model_path, exist_ok=True) 601 | Utils.download_file(url, save_path) 602 | 603 | return save_path 604 | 605 | def print_log(*args): 606 | if os.environ.get("MZ_DEV", None) is not None: 607 | print(*args) 608 | 609 | def modelscope_download_model(model_type, model_name, only_get_path=False): 610 | if model_type not in modelscope_models_map: 611 | if only_get_path: 612 | return None 613 | raise ValueError(f"模型类型 {model_type} 不支持") 614 | 615 | if model_name not in modelscope_models_map[model_type]: 616 | if only_get_path: 617 | return None 618 | error_info = "魔搭可选模型名称列表:\n" 619 | for key in modelscope_models_map[model_type].keys(): 620 | error_info += f"> {key}\n" 621 | raise ValueError(error_info) 622 | 623 | model_info = modelscope_models_map[model_type][model_name] 624 | url = model_info["url"] 625 | output = model_info["output"] 626 | save_path = os.path.abspath( 627 | os.path.join(Utils.get_models_path(), output)) 628 | if not os.path.exists(save_path): 629 | if only_get_path: 630 | return None 631 | save_path = Utils.download_file(url, save_path) 632 | return save_path 633 | 634 | def progress_bar(steps): 635 | class pb: 636 | def __init__(self, steps): 637 | self.steps = steps 638 | self.pbar = comfy.utils.ProgressBar(steps) 639 | 640 | def update(self, step, total_steps, pil_img): 641 | if pil_img is None: 642 | self.pbar.update(step, total_steps) 643 | 644 | else: 645 | if pil_img.mode != "RGB": 646 | pil_img = pil_img.convert("RGB") 647 | self.pbar.update_absolute( 648 | step, total_steps, ("JPEG", pil_img, 512)) 649 | 650 | return pb(steps) 651 | 652 | def split_en_to_zh(text: str): 653 | if text.find("(") != -1 and text.find(")") != -1: 654 | sentences = [ 655 | "", 656 | ] 657 | for word_index in range(len(text)): 658 | if text[word_index] == "(" or text[word_index] == ")": 659 | sentences.append(str(text[word_index])) 660 | sentences.append("") 661 | else: 662 | sentences[-1] += str(text[word_index]) 663 | 664 | Utils.print_log("not_translated:", sentences) 665 | for i in range(len(sentences)): 666 | if sentences[i] != "(" and sentences[i] != ")": 667 | sentences[i] = Utils.split_en_to_zh(sentences[i]) 668 | 669 | Utils.print_log("translated:", sentences) 670 | 671 | return "".join(sentences) 672 | 673 | # 中文标点转英文标点 674 | text = text.replace("，", ",") 675 | text = text.replace("。", ".") 676 | text = text.replace("？", "?") 677 | text = text.replace("！", "!") 678 | text = text.replace("；", ";") 679 | 680 | result = [] 681 | if text.find("\n") != -1: 682 | text = text.split("\n") 683 | for t in text: 684 | if t != "": 685 | result.append(Utils.split_en_to_zh(t)) 686 | else: 687 | result.append(t) 688 | return "\n".join(result) 689 | 690 | if text.find(".") != -1: 691 | text = text.split(".") 692 | for t in text: 693 | if t != "": 694 | result.append(Utils.split_en_to_zh(t)) 695 | else: 696 | result.append(t) 697 | return ".".join(result) 698 | 699 | if text.find("?") != -1: 700 | text = text.split("?") 701 | for t in text: 702 | if t != "": 703 | result.append(Utils.split_en_to_zh(t)) 704 | else: 705 | result.append(t) 706 | return "?".join(result) 707 | 708 | if text.find("!") != -1: 709 | text = text.split("!") 710 | for t in text: 711 | if t != "": 712 | result.append(Utils.split_en_to_zh(t)) 713 | else: 714 | result.append(t) 715 | return "!".join(result) 716 | 717 | if text.find(";") != -1: 718 | text = text.split(";") 719 | for t in text: 720 | if t != "": 721 | result.append(Utils.split_en_to_zh(t)) 722 | else: 723 | result.append(t) 724 | return ";".join(result) 725 | 726 | if text.find(",") != -1: 727 | text = text.split(",") 728 | for t in text: 729 | if t != "": 730 | result.append(Utils.split_en_to_zh(t)) 731 | else: 732 | result.append(t) 733 | return ",".join(result) 734 | 735 | if text.find(":") != -1: 736 | text = text.split(":") 737 | for t in text: 738 | if t != "": 739 | result.append(Utils.split_en_to_zh(t)) 740 | else: 741 | result.append(t) 742 | return ":".join(result) 743 | 744 | # 如果是纯数字,不翻译 745 | if text.isdigit() or text.replace(".", "").isdigit() or text.replace(" ", "").isdigit() or text.replace("-", "").isdigit(): 746 | return text 747 | 748 | return Utils.en2zh(text) 749 | 750 | def to_debug_prompt(p): 751 | if p is None: 752 | return "" 753 | zh = Utils.en2zh(p) 754 | if p == zh: 755 | return p 756 | zh = Utils.split_en_to_zh(p) 757 | p = p.strip() 758 | return f""" 759 | 原文: 760 | {p} 761 | 762 | 中文翻译: 763 | {zh} 764 | """ 765 | 766 | def get_gguf_files(): 767 | gguf_dir = Utils.get_gguf_models_path() 768 | if not os.path.exists(gguf_dir): 769 | os.makedirs(gguf_dir) 770 | gguf_files = [] 771 | # walk gguf_dir 772 | for root, dirs, files in os.walk(gguf_dir): 773 | for file in files: 774 | if file.endswith(".gguf"): 775 | gguf_files.append( 776 | os.path.relpath(os.path.join(root, file), gguf_dir)) 777 | 778 | return gguf_files 779 | 780 | def get_comfyui_models_path(): 781 | return folder_paths.models_dir 782 | 783 | def download_model(model_info, only_get_path=False): 784 | 785 | url = model_info["url"] 786 | output = model_info["output"] 787 | save_path = os.path.abspath( 788 | os.path.join(Utils.get_comfyui_models_path(), output)) 789 | if not os.path.exists(save_path): 790 | if only_get_path: 791 | return None 792 | save_path = Utils.download_file(url, save_path) 793 | return save_path 794 | 795 | def file_hash(file_path, hash_method): 796 | if not os.path.isfile(file_path): 797 | return '' 798 | h = hash_method() 799 | with open(file_path, 'rb') as f: 800 | while b := f.read(8192): 801 | h.update(b) 802 | return h.hexdigest() 803 | 804 | def get_cache_by_local(key): 805 | try: 806 | cache_json_file = os.path.join( 807 | Utils.get_models_path(), f"caches.json") 808 | 809 | if not os.path.exists(cache_json_file): 810 | return None 811 | 812 | with open(cache_json_file, "r", encoding="utf-8") as f: 813 | cache_json = json.load(f) 814 | return cache_json.get(key, None) 815 | except: 816 | return None 817 | 818 | def set_cache_by_local(key, value): 819 | try: 820 | cache_json_file = os.path.join( 821 | Utils.get_models_path(), f"caches.json") 822 | 823 | if not os.path.exists(cache_json_file): 824 | cache_json = {} 825 | else: 826 | with open(cache_json_file, "r", encoding="utf-8") as f: 827 | cache_json = json.load(f) 828 | 829 | cache_json[key] = value 830 | 831 | with open(cache_json_file, "w", encoding="utf-8") as f: 832 | json.dump(cache_json, f, indent=4) 833 | except: 834 | pass 835 | 836 | def file_sha256(file_path): 837 | # 获取文件的更新时间 838 | file_stat = os.stat(file_path) 839 | file_mtime = file_stat.st_mtime 840 | file_size = file_stat.st_size 841 | cache_key = f"{file_path}_{file_mtime}_{file_size}" 842 | cache_value = Utils.get_cache_by_local(cache_key) 843 | if cache_value is not None: 844 | return cache_value 845 | 846 | sha256 = Utils.file_hash(file_path, hashlib.sha256) 847 | Utils.set_cache_by_local(cache_key, sha256) 848 | return sha256 849 | 850 | def get_auto_model_fullpath(model_name): 851 | fullpath = Utils.cache_get(f"get_auto_model_fullpath_{model_name}") 852 | Utils.print_log(f"get_auto_model_fullpath_{model_name} => {fullpath}") 853 | if fullpath is not None: 854 | if os.path.exists(fullpath): 855 | return fullpath 856 | 857 | find_paths = [] 858 | target_sha256 = "" 859 | file_path = "" 860 | download_url = "" 861 | 862 | MODEL_ZOO = Utils.get_model_zoo() 863 | for model in MODEL_ZOO: 864 | if model["model"] == model_name: 865 | find_paths = model["find_path"] 866 | target_sha256 = model["SHA256"] 867 | file_path = model["file_path"] 868 | download_url = model["url"] 869 | break 870 | 871 | if target_sha256 == "": 872 | raise ValueError(f"Model {model_name} not found in MODEL_ZOO") 873 | 874 | if os.path.exists(file_path): 875 | if Utils.file_sha256(file_path) != target_sha256: 876 | print(f"Model {model_name} file hash not match...") 877 | return file_path 878 | 879 | for find_path in find_paths: 880 | find_fullpath = os.path.join( 881 | Utils.get_comfyui_models_path(), find_path) 882 | 883 | if os.path.exists(find_fullpath): 884 | for root, dirs, files in os.walk(find_fullpath): 885 | for file in files: 886 | if target_sha256 == Utils.file_sha256(os.path.join(root, file)): 887 | Utils.cache_set( 888 | f"get_auto_model_fullpath_{model_name}", os.path.join(root, file)) 889 | return os.path.join(root, file) 890 | else: 891 | Utils.print_log( 892 | f"Model {os.path.join(root, file)} file hash not match, {target_sha256} != {Utils.file_sha256(os.path.join(root, file))}") 893 | 894 | result = Utils.download_model( 895 | {"url": download_url, "output": file_path}) 896 | Utils.cache_set(f"get_auto_model_fullpath_{model_name}", result) 897 | return result 898 | 899 | def testDownloadSpeed(url): 900 | try: 901 | print(f"Testing download speed for {url}") 902 | start = time.time() 903 | # 下载2M数据 904 | headers = {"Range": "bytes=0-2097151"} 905 | _ = requests.get(url, headers=headers, timeout=5) 906 | end = time.time() 907 | print( 908 | f"Download speed: {round(5.00 / (float(end) - float(start)) / 1024, 2)} KB/s") 909 | return float(end) - float(start) < 4 910 | except Exception as e: 911 | print(f"Test download speed failed: {e}") 912 | return False 913 | 914 | def get_model_zoo(tags_filter=None): 915 | source_model_zoo_file = os.path.join( 916 | os.path.dirname(__file__), "configs", "model_zoo.json") 917 | source_model_zoo_json = [] 918 | try: 919 | with open(source_model_zoo_file, "r", encoding="utf-8") as f: 920 | source_model_zoo_json = json.load(f) 921 | except: 922 | pass 923 | 924 | # Utils.print_log(f"source_model_zoo_json: {json.dumps(source_model_zoo_json, indent=4)}") 925 | if tags_filter is not None: 926 | source_model_zoo_json = [ 927 | m for m in source_model_zoo_json if tags_filter in m["tags"]] 928 | 929 | return source_model_zoo_json 930 | 931 | 932 | 933 | modelscope_models_map = { 934 | 935 | } 936 | --------------------------------------------------------------------------------