Paul Borne--Pons, Mikolaj Czerkawski,Rosalie Martin, 5 | Romain Rouffet
6 | 7 | 8 | 9 | [](https://paulbornep.github.io/mesa-terrain/) 10 | [](https://arxiv.org/abs/2504.07210) 11 | [](https://www.huggingface.co/NewtNewt/MESA) 12 | [](https://www.huggingface.co/Major-TOM) 13 | [](https://huggingface.co/spaces/mikonvergence/MESA) 14 |
15 |
16 |
17 | MESA is a novel generative model based on latent denoising diffusion capable of generating 2.5D representations of terrain based on the text prompt conditioning supplied via natural language. The model produces two co-registered modalities of optical and depth maps.
18 |
19 | 
57 |
58 | Alternatively, you can download and use the Gradio demo from the HF page.
59 |
60 | ## Citation
61 |
62 | ```latex
63 | @inproceedings{mesa2025,
64 | title={MESA: Text-Driven Terrain Generation Using Latent Diffusion and Global Copernicus Data},
65 | author={Paul Borne--Pons and Mikolaj Czerkawski and Rosalie Martin and Romain Rouffet},
66 | year={2025},
67 | booktitle={MORSE Workshop at CVPR 2025},
68 | eprint={2504.07210},
69 | url={https://arxiv.org/abs/2504.07210},}
70 | ```
71 | ## Acknowledgements
72 |
73 | This implementation builds upon Hugging Face’s [Diffusers](https://github.com/huggingface/diffusers) library. We also acknowledge [Gradio](https://www.gradio.app/) for providing an easy-to-use interface that allowed us to create the inference demos for our models.
74 |
75 | This model is the product of a collaboration between [Φ-lab, European Space Agency (ESA)](https://philab.esa.int/) and the [Adobe Research (Paris, France)](https://research.adobe.com/careers/paris/).
76 |
--------------------------------------------------------------------------------
/pipeline_terrain.py:
--------------------------------------------------------------------------------
1 | ###########################################################################
2 | # References:
3 | # https://github.com/huggingface/diffusers/
4 | ###########################################################################
5 | import inspect
6 | from typing import Any, Callable, Dict, List, Optional, Union
7 |
8 | import torch
9 | from packaging import version
10 | from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
11 |
12 | from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
13 | from diffusers.configuration_utils import FrozenDict
14 | from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
15 | from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
16 | from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
17 | from diffusers.models.lora import adjust_lora_scale_text_encoder
18 | from diffusers.schedulers import KarrasDiffusionSchedulers
19 | from diffusers.utils import (
20 | USE_PEFT_BACKEND,
21 | deprecate,
22 | is_torch_xla_available,
23 | logging,
24 | replace_example_docstring,
25 | scale_lora_layers,
26 | unscale_lora_layers,
27 | )
28 | from diffusers.utils.torch_utils import randn_tensor
29 | from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
30 | from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
31 | from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
32 |
33 |
34 | if is_torch_xla_available():
35 | import torch_xla.core.xla_model as xm
36 |
37 | XLA_AVAILABLE = True
38 | else:
39 | XLA_AVAILABLE = False
40 |
41 | logger = logging.get_logger(__name__) # pylint: disable=invalid-name
42 |
43 | EXAMPLE_DOC_STRING = """
44 | Examples:
45 | ```py
46 | >>> import torch
47 | >>> from diffusers import StableDiffusionPipeline
48 |
49 | >>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
50 | >>> pipe = pipe.to("cuda")
51 |
52 | >>> prompt = "a photo of an astronaut riding a horse on mars"
53 | >>> image = pipe(prompt).images[0]
54 | ```
55 | """
56 |
57 |
58 | def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
59 | """
60 | Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
61 | Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
62 | """
63 | std_text = noise_pred_text.std(
64 | dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
65 | std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
66 | # rescale the results from guidance (fixes overexposure)
67 | noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
68 | # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
69 | noise_cfg = guidance_rescale * noise_pred_rescaled + \
70 | (1 - guidance_rescale) * noise_cfg
71 | return noise_cfg
72 |
73 |
74 | def retrieve_timesteps(
75 | scheduler,
76 | num_inference_steps: Optional[int] = None,
77 | device: Optional[Union[str, torch.device]] = None,
78 | timesteps: Optional[List[int]] = None,
79 | sigmas: Optional[List[float]] = None,
80 | **kwargs,
81 | ):
82 | """
83 | Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
84 | custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
85 |
86 | Args:
87 | scheduler (`SchedulerMixin`):
88 | The scheduler to get timesteps from.
89 | num_inference_steps (`int`):
90 | The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
91 | must be `None`.
92 | device (`str` or `torch.device`, *optional*):
93 | The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
94 | timesteps (`List[int]`, *optional*):
95 | Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
96 | `num_inference_steps` and `sigmas` must be `None`.
97 | sigmas (`List[float]`, *optional*):
98 | Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
99 | `num_inference_steps` and `timesteps` must be `None`.
100 |
101 | Returns:
102 | `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
103 | second element is the number of inference steps.
104 | """
105 | if timesteps is not None and sigmas is not None:
106 | raise ValueError(
107 | "Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
108 | if timesteps is not None:
109 | accepts_timesteps = "timesteps" in set(
110 | inspect.signature(scheduler.set_timesteps).parameters.keys())
111 | if not accepts_timesteps:
112 | raise ValueError(
113 | f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
114 | f" timestep schedules. Please check whether you are using the correct scheduler."
115 | )
116 | scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
117 | timesteps = scheduler.timesteps
118 | num_inference_steps = len(timesteps)
119 | elif sigmas is not None:
120 | accept_sigmas = "sigmas" in set(inspect.signature(
121 | scheduler.set_timesteps).parameters.keys())
122 | if not accept_sigmas:
123 | raise ValueError(
124 | f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
125 | f" sigmas schedules. Please check whether you are using the correct scheduler."
126 | )
127 | scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
128 | timesteps = scheduler.timesteps
129 | num_inference_steps = len(timesteps)
130 | else:
131 | scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
132 | timesteps = scheduler.timesteps
133 | return timesteps, num_inference_steps
134 |
135 |
136 | class TerrainDiffusionPipeline(
137 | DiffusionPipeline,
138 | StableDiffusionMixin,
139 | TextualInversionLoaderMixin,
140 | StableDiffusionLoraLoaderMixin,
141 | IPAdapterMixin,
142 | FromSingleFileMixin,
143 | ):
144 | r"""
145 | Pipeline for text-to-image generation using Stable Diffusion.
146 |
147 | This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
148 | implemented for all pipelines (downloading, saving, running on a particular device, etc.).
149 |
150 | The pipeline also inherits the following loading methods:
151 | - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
152 | - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
153 | - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
154 | - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
155 | - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
156 |
157 | Args:
158 | vae ([`AutoencoderKL`]):
159 | Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
160 | text_encoder ([`~transformers.CLIPTextModel`]):
161 | Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
162 | tokenizer ([`~transformers.CLIPTokenizer`]):
163 | A `CLIPTokenizer` to tokenize text.
164 | unet ([`UNet2DConditionModel`]):
165 | A `UNet2DConditionModel` to denoise the encoded image latents.
166 | scheduler ([`SchedulerMixin`]):
167 | A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
168 | [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
169 | safety_checker ([`StableDiffusionSafetyChecker`]):
170 | Classification module that estimates whether generated images could be considered offensive or harmful.
171 | Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
172 | about a model's potential harms.
173 | feature_extractor ([`~transformers.CLIPImageProcessor`]):
174 | A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
175 | """
176 |
177 | model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
178 | _optional_components = ["safety_checker",
179 | "feature_extractor", "image_encoder"]
180 | _exclude_from_cpu_offload = ["safety_checker"]
181 | _callback_tensor_inputs = ["latents",
182 | "prompt_embeds", "negative_prompt_embeds"]
183 |
184 | def __init__(
185 | self,
186 | vae: AutoencoderKL,
187 | text_encoder: CLIPTextModel,
188 | tokenizer: CLIPTokenizer,
189 | unet: UNet2DConditionModel,
190 | scheduler: KarrasDiffusionSchedulers,
191 | safety_checker: StableDiffusionSafetyChecker,
192 | feature_extractor: CLIPImageProcessor,
193 | image_encoder: CLIPVisionModelWithProjection = None,
194 | requires_safety_checker: bool = True,
195 | ):
196 | super().__init__()
197 |
198 | if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
199 | deprecation_message = (
200 | f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
201 | f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
202 | "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
203 | " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
204 | " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
205 | " file"
206 | )
207 | deprecate("steps_offset!=1", "1.0.0",
208 | deprecation_message, standard_warn=False)
209 | new_config = dict(scheduler.config)
210 | new_config["steps_offset"] = 1
211 | scheduler._internal_dict = FrozenDict(new_config)
212 |
213 | if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
214 | deprecation_message = (
215 | f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
216 | " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
217 | " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
218 | " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
219 | " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
220 | )
221 | deprecate("clip_sample not set", "1.0.0",
222 | deprecation_message, standard_warn=False)
223 | new_config = dict(scheduler.config)
224 | new_config["clip_sample"] = False
225 | scheduler._internal_dict = FrozenDict(new_config)
226 |
227 | if safety_checker is None and requires_safety_checker:
228 | logger.warning(
229 | f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
230 | " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
231 | " results in services or applications open to the public. Both the diffusers team and Hugging Face"
232 | " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
233 | " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
234 | " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
235 | )
236 |
237 | if safety_checker is not None and feature_extractor is None:
238 | raise ValueError(
239 | "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
240 | " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
241 | )
242 |
243 | is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
244 | version.parse(unet.config._diffusers_version).base_version
245 | ) < version.parse("0.9.0.dev0")
246 | is_unet_sample_size_less_64 = hasattr(
247 | unet.config, "sample_size") and unet.config.sample_size < 64
248 | if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
249 | deprecation_message = (
250 | "The configuration file of the unet has set the default `sample_size` to smaller than"
251 | " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
252 | " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
253 | " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
254 | " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
255 | " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
256 | " in the config might lead to incorrect results in future versions. If you have downloaded this"
257 | " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
258 | " the `unet/config.json` file"
259 | )
260 | deprecate("sample_size<64", "1.0.0",
261 | deprecation_message, standard_warn=False)
262 | new_config = dict(unet.config)
263 | new_config["sample_size"] = 64
264 | unet._internal_dict = FrozenDict(new_config)
265 |
266 | self.register_modules(
267 | vae=vae,
268 | text_encoder=text_encoder,
269 | tokenizer=tokenizer,
270 | unet=unet,
271 | scheduler=scheduler,
272 | safety_checker=safety_checker,
273 | feature_extractor=feature_extractor,
274 | image_encoder=image_encoder,
275 | )
276 | self.vae_scale_factor = 2 ** (
277 | len(self.vae.config.block_out_channels) - 1)
278 | self.image_processor = VaeImageProcessor(
279 | vae_scale_factor=self.vae_scale_factor)
280 | self.register_to_config(
281 | requires_safety_checker=requires_safety_checker)
282 |
283 | def _encode_prompt(
284 | self,
285 | prompt,
286 | device,
287 | num_images_per_prompt,
288 | do_classifier_free_guidance,
289 | negative_prompt=None,
290 | prompt_embeds: Optional[torch.Tensor] = None,
291 | negative_prompt_embeds: Optional[torch.Tensor] = None,
292 | lora_scale: Optional[float] = None,
293 | **kwargs,
294 | ):
295 | deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
296 | deprecate("_encode_prompt()", "1.0.0",
297 | deprecation_message, standard_warn=False)
298 |
299 | prompt_embeds_tuple = self.encode_prompt(
300 | prompt=prompt,
301 | device=device,
302 | num_images_per_prompt=num_images_per_prompt,
303 | do_classifier_free_guidance=do_classifier_free_guidance,
304 | negative_prompt=negative_prompt,
305 | prompt_embeds=prompt_embeds,
306 | negative_prompt_embeds=negative_prompt_embeds,
307 | lora_scale=lora_scale,
308 | **kwargs,
309 | )
310 |
311 | # concatenate for backwards comp
312 | prompt_embeds = torch.cat(
313 | [prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
314 |
315 | return prompt_embeds
316 |
317 | def encode_prompt(
318 | self,
319 | prompt,
320 | device,
321 | num_images_per_prompt,
322 | do_classifier_free_guidance,
323 | negative_prompt=None,
324 | prompt_embeds: Optional[torch.Tensor] = None,
325 | negative_prompt_embeds: Optional[torch.Tensor] = None,
326 | lora_scale: Optional[float] = None,
327 | clip_skip: Optional[int] = None,
328 | ):
329 | r"""
330 | Encodes the prompt into text encoder hidden states.
331 |
332 | Args:
333 | prompt (`str` or `List[str]`, *optional*):
334 | prompt to be encoded
335 | device: (`torch.device`):
336 | torch device
337 | num_images_per_prompt (`int`):
338 | number of images that should be generated per prompt
339 | do_classifier_free_guidance (`bool`):
340 | whether to use classifier free guidance or not
341 | negative_prompt (`str` or `List[str]`, *optional*):
342 | The prompt or prompts not to guide the image generation. If not defined, one has to pass
343 | `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
344 | less than `1`).
345 | prompt_embeds (`torch.Tensor`, *optional*):
346 | Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
347 | provided, text embeddings will be generated from `prompt` input argument.
348 | negative_prompt_embeds (`torch.Tensor`, *optional*):
349 | Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
350 | weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
351 | argument.
352 | lora_scale (`float`, *optional*):
353 | A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
354 | clip_skip (`int`, *optional*):
355 | Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
356 | the output of the pre-final layer will be used for computing the prompt embeddings.
357 | """
358 | # set lora scale so that monkey patched LoRA
359 | # function of text encoder can correctly access it
360 | if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
361 | self._lora_scale = lora_scale
362 |
363 | # dynamically adjust the LoRA scale
364 | if not USE_PEFT_BACKEND:
365 | adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
366 | else:
367 | scale_lora_layers(self.text_encoder, lora_scale)
368 |
369 | if prompt is not None and isinstance(prompt, str):
370 | batch_size = 1
371 | elif prompt is not None and isinstance(prompt, list):
372 | batch_size = len(prompt)
373 | else:
374 | batch_size = prompt_embeds.shape[0]
375 |
376 | if prompt_embeds is None:
377 | # textual inversion: process multi-vector tokens if necessary
378 | if isinstance(self, TextualInversionLoaderMixin):
379 | prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
380 |
381 | text_inputs = self.tokenizer(
382 | prompt,
383 | padding="max_length",
384 | max_length=self.tokenizer.model_max_length,
385 | truncation=True,
386 | return_tensors="pt",
387 | )
388 | text_input_ids = text_inputs.input_ids
389 | untruncated_ids = self.tokenizer(
390 | prompt, padding="longest", return_tensors="pt").input_ids
391 |
392 | if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
393 | text_input_ids, untruncated_ids
394 | ):
395 | removed_text = self.tokenizer.batch_decode(
396 | untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
397 | )
398 | logger.warning(
399 | "The following part of your input was truncated because CLIP can only handle sequences up to"
400 | f" {self.tokenizer.model_max_length} tokens: {removed_text}"
401 | )
402 |
403 | if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
404 | attention_mask = text_inputs.attention_mask.to(device)
405 | else:
406 | attention_mask = None
407 |
408 | if clip_skip is None:
409 | prompt_embeds = self.text_encoder(
410 | text_input_ids.to(device), attention_mask=attention_mask)
411 | prompt_embeds = prompt_embeds[0]
412 | else:
413 | prompt_embeds = self.text_encoder(
414 | text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
415 | )
416 | # Access the `hidden_states` first, that contains a tuple of
417 | # all the hidden states from the encoder layers. Then index into
418 | # the tuple to access the hidden states from the desired layer.
419 | prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
420 | # We also need to apply the final LayerNorm here to not mess with the
421 | # representations. The `last_hidden_states` that we typically use for
422 | # obtaining the final prompt representations passes through the LayerNorm
423 | # layer.
424 | prompt_embeds = self.text_encoder.text_model.final_layer_norm(
425 | prompt_embeds)
426 |
427 | if self.text_encoder is not None:
428 | prompt_embeds_dtype = self.text_encoder.dtype
429 | elif self.unet is not None:
430 | prompt_embeds_dtype = self.unet.dtype
431 | else:
432 | prompt_embeds_dtype = prompt_embeds.dtype
433 |
434 | prompt_embeds = prompt_embeds.to(
435 | dtype=prompt_embeds_dtype, device=device)
436 |
437 | bs_embed, seq_len, _ = prompt_embeds.shape
438 | # duplicate text embeddings for each generation per prompt, using mps friendly method
439 | prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
440 | prompt_embeds = prompt_embeds.view(
441 | bs_embed * num_images_per_prompt, seq_len, -1)
442 |
443 | # get unconditional embeddings for classifier free guidance
444 | if do_classifier_free_guidance and negative_prompt_embeds is None:
445 | uncond_tokens: List[str]
446 | if negative_prompt is None:
447 | uncond_tokens = [""] * batch_size
448 | elif prompt is not None and type(prompt) is not type(negative_prompt):
449 | raise TypeError(
450 | f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
451 | f" {type(prompt)}."
452 | )
453 | elif isinstance(negative_prompt, str):
454 | uncond_tokens = [negative_prompt]
455 | elif batch_size != len(negative_prompt):
456 | raise ValueError(
457 | f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
458 | f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
459 | " the batch size of `prompt`."
460 | )
461 | else:
462 | uncond_tokens = negative_prompt
463 |
464 | # textual inversion: process multi-vector tokens if necessary
465 | if isinstance(self, TextualInversionLoaderMixin):
466 | uncond_tokens = self.maybe_convert_prompt(
467 | uncond_tokens, self.tokenizer)
468 |
469 | max_length = prompt_embeds.shape[1]
470 | uncond_input = self.tokenizer(
471 | uncond_tokens,
472 | padding="max_length",
473 | max_length=max_length,
474 | truncation=True,
475 | return_tensors="pt",
476 | )
477 |
478 | if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
479 | attention_mask = uncond_input.attention_mask.to(device)
480 | else:
481 | attention_mask = None
482 |
483 | negative_prompt_embeds = self.text_encoder(
484 | uncond_input.input_ids.to(device),
485 | attention_mask=attention_mask,
486 | )
487 | negative_prompt_embeds = negative_prompt_embeds[0]
488 |
489 | if do_classifier_free_guidance:
490 | # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
491 | seq_len = negative_prompt_embeds.shape[1]
492 |
493 | negative_prompt_embeds = negative_prompt_embeds.to(
494 | dtype=prompt_embeds_dtype, device=device)
495 |
496 | negative_prompt_embeds = negative_prompt_embeds.repeat(
497 | 1, num_images_per_prompt, 1)
498 | negative_prompt_embeds = negative_prompt_embeds.view(
499 | batch_size * num_images_per_prompt, seq_len, -1)
500 |
501 | if self.text_encoder is not None:
502 | if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
503 | # Retrieve the original scale by scaling back the LoRA layers
504 | unscale_lora_layers(self.text_encoder, lora_scale)
505 |
506 | return prompt_embeds, negative_prompt_embeds
507 |
508 | def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
509 | dtype = next(self.image_encoder.parameters()).dtype
510 |
511 | if not isinstance(image, torch.Tensor):
512 | image = self.feature_extractor(
513 | image, return_tensors="pt").pixel_values
514 |
515 | image = image.to(device=device, dtype=dtype)
516 | if output_hidden_states:
517 | image_enc_hidden_states = self.image_encoder(
518 | image, output_hidden_states=True).hidden_states[-2]
519 | image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(
520 | num_images_per_prompt, dim=0)
521 | uncond_image_enc_hidden_states = self.image_encoder(
522 | torch.zeros_like(image), output_hidden_states=True
523 | ).hidden_states[-2]
524 | uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
525 | num_images_per_prompt, dim=0
526 | )
527 | return image_enc_hidden_states, uncond_image_enc_hidden_states
528 | else:
529 | image_embeds = self.image_encoder(image).image_embeds
530 | image_embeds = image_embeds.repeat_interleave(
531 | num_images_per_prompt, dim=0)
532 | uncond_image_embeds = torch.zeros_like(image_embeds)
533 |
534 | return image_embeds, uncond_image_embeds
535 |
536 | def prepare_ip_adapter_image_embeds(
537 | self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
538 | ):
539 | image_embeds = []
540 | if do_classifier_free_guidance:
541 | negative_image_embeds = []
542 | if ip_adapter_image_embeds is None:
543 | if not isinstance(ip_adapter_image, list):
544 | ip_adapter_image = [ip_adapter_image]
545 |
546 | if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
547 | raise ValueError(
548 | f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
549 | )
550 |
551 | for single_ip_adapter_image, image_proj_layer in zip(
552 | ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
553 | ):
554 | output_hidden_state = not isinstance(
555 | image_proj_layer, ImageProjection)
556 | single_image_embeds, single_negative_image_embeds = self.encode_image(
557 | single_ip_adapter_image, device, 1, output_hidden_state
558 | )
559 |
560 | image_embeds.append(single_image_embeds[None, :])
561 | if do_classifier_free_guidance:
562 | negative_image_embeds.append(
563 | single_negative_image_embeds[None, :])
564 | else:
565 | for single_image_embeds in ip_adapter_image_embeds:
566 | if do_classifier_free_guidance:
567 | single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(
568 | 2)
569 | negative_image_embeds.append(single_negative_image_embeds)
570 | image_embeds.append(single_image_embeds)
571 |
572 | ip_adapter_image_embeds = []
573 | for i, single_image_embeds in enumerate(image_embeds):
574 | single_image_embeds = torch.cat(
575 | [single_image_embeds] * num_images_per_prompt, dim=0)
576 | if do_classifier_free_guidance:
577 | single_negative_image_embeds = torch.cat(
578 | [negative_image_embeds[i]] * num_images_per_prompt, dim=0)
579 | single_image_embeds = torch.cat(
580 | [single_negative_image_embeds, single_image_embeds], dim=0)
581 |
582 | single_image_embeds = single_image_embeds.to(device=device)
583 | ip_adapter_image_embeds.append(single_image_embeds)
584 |
585 | return ip_adapter_image_embeds
586 |
587 | def decode_latents(self, latents):
588 | deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
589 | deprecate("decode_latents", "1.0.0",
590 | deprecation_message, standard_warn=False)
591 |
592 | latents = 1 / self.vae.config.scaling_factor * latents
593 | image = self.vae.decode(latents, return_dict=False)[0]
594 | image = (image / 2 + 0.5).clamp(0, 1)
595 | # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
596 | image = image.cpu().permute(0, 2, 3, 1).float().numpy()
597 | return image
598 |
599 | def prepare_extra_step_kwargs(self, generator, eta):
600 | # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
601 | # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
602 | # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
603 | # and should be between [0, 1]
604 |
605 | accepts_eta = "eta" in set(inspect.signature(
606 | self.scheduler.step).parameters.keys())
607 | extra_step_kwargs = {}
608 | if accepts_eta:
609 | extra_step_kwargs["eta"] = eta
610 |
611 | # check if the scheduler accepts generator
612 | accepts_generator = "generator" in set(
613 | inspect.signature(self.scheduler.step).parameters.keys())
614 | if accepts_generator:
615 | extra_step_kwargs["generator"] = generator
616 | return extra_step_kwargs
617 |
618 | def check_inputs(
619 | self,
620 | prompt,
621 | height,
622 | width,
623 | callback_steps,
624 | negative_prompt=None,
625 | prompt_embeds=None,
626 | negative_prompt_embeds=None,
627 | ip_adapter_image=None,
628 | ip_adapter_image_embeds=None,
629 | callback_on_step_end_tensor_inputs=None,
630 | ):
631 | if height % 8 != 0 or width % 8 != 0:
632 | raise ValueError(
633 | f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
634 |
635 | if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
636 | raise ValueError(
637 | f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
638 | f" {type(callback_steps)}."
639 | )
640 | if callback_on_step_end_tensor_inputs is not None and not all(
641 | k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
642 | ):
643 | raise ValueError(
644 | f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
645 | )
646 |
647 | if prompt is not None and prompt_embeds is not None:
648 | raise ValueError(
649 | f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
650 | " only forward one of the two."
651 | )
652 | elif prompt is None and prompt_embeds is None:
653 | raise ValueError(
654 | "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
655 | )
656 | elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
657 | raise ValueError(
658 | f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
659 |
660 | if negative_prompt is not None and negative_prompt_embeds is not None:
661 | raise ValueError(
662 | f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
663 | f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
664 | )
665 |
666 | if prompt_embeds is not None and negative_prompt_embeds is not None:
667 | if prompt_embeds.shape != negative_prompt_embeds.shape:
668 | raise ValueError(
669 | "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
670 | f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
671 | f" {negative_prompt_embeds.shape}."
672 | )
673 |
674 | if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
675 | raise ValueError(
676 | "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
677 | )
678 |
679 | if ip_adapter_image_embeds is not None:
680 | if not isinstance(ip_adapter_image_embeds, list):
681 | raise ValueError(
682 | f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
683 | )
684 | elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
685 | raise ValueError(
686 | f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
687 | )
688 |
689 | def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
690 | shape = (
691 | batch_size,
692 | num_channels_latents,
693 | int(height) // self.vae_scale_factor,
694 | int(width) // self.vae_scale_factor,
695 | )
696 | if isinstance(generator, list) and len(generator) != batch_size:
697 | raise ValueError(
698 | f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
699 | f" size of {batch_size}. Make sure the batch size matches the length of the generators."
700 | )
701 |
702 | if latents is None:
703 | latents = randn_tensor(
704 | shape, generator=generator, device=device, dtype=dtype)
705 | else:
706 | latents = latents.to(device)
707 |
708 | # scale the initial noise by the standard deviation required by the scheduler
709 | latents = latents * self.scheduler.init_noise_sigma
710 | return latents
711 |
712 | # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
713 | def get_guidance_scale_embedding(
714 | self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32
715 | ) -> torch.Tensor:
716 | """
717 | See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
718 |
719 | Args:
720 | w (`torch.Tensor`):
721 | Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
722 | embedding_dim (`int`, *optional*, defaults to 512):
723 | Dimension of the embeddings to generate.
724 | dtype (`torch.dtype`, *optional*, defaults to `torch.float32`):
725 | Data type of the generated embeddings.
726 |
727 | Returns:
728 | `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
729 | """
730 | assert len(w.shape) == 1
731 | w = w * 1000.0
732 |
733 | half_dim = embedding_dim // 2
734 | emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
735 | emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
736 | emb = w.to(dtype)[:, None] * emb[None, :]
737 | emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
738 | if embedding_dim % 2 == 1: # zero pad
739 | emb = torch.nn.functional.pad(emb, (0, 1))
740 | assert emb.shape == (w.shape[0], embedding_dim)
741 | return emb
742 |
743 | @property
744 | def guidance_scale(self):
745 | return self._guidance_scale
746 |
747 | @property
748 | def guidance_rescale(self):
749 | return self._guidance_rescale
750 |
751 | @property
752 | def clip_skip(self):
753 | return self._clip_skip
754 |
755 | # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
756 | # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
757 | # corresponds to doing no classifier free guidance.
758 | @property
759 | def do_classifier_free_guidance(self):
760 | return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
761 |
762 | @property
763 | def cross_attention_kwargs(self):
764 | return self._cross_attention_kwargs
765 |
766 | @property
767 | def num_timesteps(self):
768 | return self._num_timesteps
769 |
770 | @property
771 | def interrupt(self):
772 | return self._interrupt
773 |
774 | def decode_rgbd(self, latents, generator, output_type="np"):
775 | dem_latents = latents[:, 4:, :, :]
776 | img_latents = latents[:, :4, :, :]
777 | image = self.vae.decode(img_latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
778 | 0
779 | ]
780 | dem = self.vae.decode(dem_latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
781 | 0
782 | ]
783 | do_denormalize = [True] * image.shape[0]
784 | image = self.image_processor.postprocess(
785 | image, output_type=output_type, do_denormalize=do_denormalize)
786 | dem = self.image_processor.postprocess(
787 | dem, output_type=output_type, do_denormalize=do_denormalize)
788 | return image, dem
789 |
790 | @torch.no_grad()
791 | @replace_example_docstring(EXAMPLE_DOC_STRING)
792 | def __call__(
793 | self,
794 | prompt: Union[str, List[str]] = None,
795 | height: Optional[int] = None,
796 | width: Optional[int] = None,
797 | num_inference_steps: int = 50,
798 | timesteps: List[int] = None,
799 | sigmas: List[float] = None,
800 | guidance_scale: float = 7.5,
801 | negative_prompt: Optional[Union[str, List[str]]] = None,
802 | num_images_per_prompt: Optional[int] = 1,
803 | eta: float = 0.0,
804 | generator: Optional[Union[torch.Generator,
805 | List[torch.Generator]]] = None,
806 | latents: Optional[torch.Tensor] = None,
807 | prompt_embeds: Optional[torch.Tensor] = None,
808 | negative_prompt_embeds: Optional[torch.Tensor] = None,
809 | ip_adapter_image: Optional[PipelineImageInput] = None,
810 | ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
811 | output_type: Optional[str] = "np",
812 | return_dict: bool = True,
813 | cross_attention_kwargs: Optional[Dict[str, Any]] = None,
814 | guidance_rescale: float = 0.0,
815 | clip_skip: Optional[int] = None,
816 | callback_on_step_end: Optional[
817 | Union[Callable[[int, int, Dict], None],
818 | PipelineCallback, MultiPipelineCallbacks]
819 | ] = None,
820 | callback_on_step_end_tensor_inputs: List[str] = ["latents"],
821 | **kwargs,
822 | ):
823 | r"""
824 | The call function to the pipeline for generation.
825 |
826 | Args:
827 | prompt (`str` or `List[str]`, *optional*):
828 | The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
829 | height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
830 | The height in pixels of the generated image.
831 | width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
832 | The width in pixels of the generated image.
833 | num_inference_steps (`int`, *optional*, defaults to 50):
834 | The number of denoising steps. More denoising steps usually lead to a higher quality image at the
835 | expense of slower inference.
836 | timesteps (`List[int]`, *optional*):
837 | Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
838 | in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
839 | passed will be used. Must be in descending order.
840 | sigmas (`List[float]`, *optional*):
841 | Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
842 | their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
843 | will be used.
844 | guidance_scale (`float`, *optional*, defaults to 7.5):
845 | A higher guidance scale value encourages the model to generate images closely linked to the text
846 | `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
847 | negative_prompt (`str` or `List[str]`, *optional*):
848 | The prompt or prompts to guide what to not include in image generation. If not defined, you need to
849 | pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
850 | num_images_per_prompt (`int`, *optional*, defaults to 1):
851 | The number of images to generate per prompt.
852 | eta (`float`, *optional*, defaults to 0.0):
853 | Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
854 | to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
855 | generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
856 | A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
857 | generation deterministic.
858 | latents (`torch.Tensor`, *optional*):
859 | Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
860 | generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
861 | tensor is generated by sampling using the supplied random `generator`.
862 | prompt_embeds (`torch.Tensor`, *optional*):
863 | Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
864 | provided, text embeddings are generated from the `prompt` input argument.
865 | negative_prompt_embeds (`torch.Tensor`, *optional*):
866 | Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
867 | not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
868 | ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
869 | ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
870 | Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
871 | IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
872 | contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
873 | provided, embeddings are computed from the `ip_adapter_image` input argument.
874 | output_type (`str`, *optional*, defaults to `"pil"`):
875 | The output format of the generated image. Choose between `PIL.Image` or `np.array`.
876 | return_dict (`bool`, *optional*, defaults to `True`):
877 | Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
878 | plain tuple.
879 | cross_attention_kwargs (`dict`, *optional*):
880 | A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
881 | [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
882 | guidance_rescale (`float`, *optional*, defaults to 0.0):
883 | Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
884 | Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
885 | using zero terminal SNR.
886 | clip_skip (`int`, *optional*):
887 | Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
888 | the output of the pre-final layer will be used for computing the prompt embeddings.
889 | callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
890 | A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
891 | each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
892 | DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
893 | list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
894 | callback_on_step_end_tensor_inputs (`List`, *optional*):
895 | The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
896 | will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
897 | `._callback_tensor_inputs` attribute of your pipeline class.
898 |
899 | Examples:
900 |
901 | Returns:
902 | [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
903 | If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
904 | otherwise a `tuple` is returned where the first element is a list with the generated images and the
905 | second element is a list of `bool`s indicating whether the corresponding generated image contains
906 | "not-safe-for-work" (nsfw) content.
907 | """
908 |
909 | callback = kwargs.pop("callback", None)
910 | callback_steps = kwargs.pop("callback_steps", None)
911 |
912 | if callback is not None:
913 | deprecate(
914 | "callback",
915 | "1.0.0",
916 | "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
917 | )
918 | if callback_steps is not None:
919 | deprecate(
920 | "callback_steps",
921 | "1.0.0",
922 | "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
923 | )
924 |
925 | if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
926 | callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
927 |
928 | # 0. Default height and width to unet
929 | height = height or self.unet.config.sample_size * self.vae_scale_factor
930 | width = width or self.unet.config.sample_size * self.vae_scale_factor
931 | # to deal with lora scaling and other possible forward hooks
932 |
933 | # 1. Check inputs. Raise error if not correct
934 | self.check_inputs(
935 | prompt,
936 | height,
937 | width,
938 | callback_steps,
939 | negative_prompt,
940 | prompt_embeds,
941 | negative_prompt_embeds,
942 | ip_adapter_image,
943 | ip_adapter_image_embeds,
944 | callback_on_step_end_tensor_inputs,
945 | )
946 |
947 | self._guidance_scale = guidance_scale
948 | self._guidance_rescale = guidance_rescale
949 | self._clip_skip = clip_skip
950 | self._cross_attention_kwargs = cross_attention_kwargs
951 | self._interrupt = False
952 |
953 | # 2. Define call parameters
954 | if prompt is not None and isinstance(prompt, str):
955 | batch_size = 1
956 | elif prompt is not None and isinstance(prompt, list):
957 | batch_size = len(prompt)
958 | else:
959 | batch_size = prompt_embeds.shape[0]
960 |
961 | device = self._execution_device
962 |
963 | # 3. Encode input prompt
964 | lora_scale = (
965 | self.cross_attention_kwargs.get(
966 | "scale", None) if self.cross_attention_kwargs is not None else None
967 | )
968 |
969 | prompt_embeds, negative_prompt_embeds = self.encode_prompt(
970 | prompt,
971 | device,
972 | num_images_per_prompt,
973 | self.do_classifier_free_guidance,
974 | negative_prompt,
975 | prompt_embeds=prompt_embeds,
976 | negative_prompt_embeds=negative_prompt_embeds,
977 | lora_scale=lora_scale,
978 | clip_skip=self.clip_skip,
979 | )
980 |
981 | # For classifier free guidance, we need to do two forward passes.
982 | # Here we concatenate the unconditional and text embeddings into a single batch
983 | # to avoid doing two forward passes
984 | if self.do_classifier_free_guidance:
985 | prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
986 |
987 | if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
988 | image_embeds = self.prepare_ip_adapter_image_embeds(
989 | ip_adapter_image,
990 | ip_adapter_image_embeds,
991 | device,
992 | batch_size * num_images_per_prompt,
993 | self.do_classifier_free_guidance,
994 | )
995 |
996 | # 4. Prepare timesteps
997 | timesteps, num_inference_steps = retrieve_timesteps(
998 | self.scheduler, num_inference_steps, device, timesteps, sigmas
999 | )
1000 |
1001 | # 5. Prepare latent variables
1002 | num_channels_latents = self.unet.config.in_channels*2
1003 | latents = self.prepare_latents(
1004 | batch_size * num_images_per_prompt,
1005 | num_channels_latents,
1006 | height,
1007 | width,
1008 | prompt_embeds.dtype,
1009 | device,
1010 | generator,
1011 | latents,
1012 | )
1013 |
1014 | # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
1015 | extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
1016 |
1017 | # 6.1 Add image embeds for IP-Adapter
1018 | added_cond_kwargs = (
1019 | {"image_embeds": image_embeds}
1020 | if (ip_adapter_image is not None or ip_adapter_image_embeds is not None)
1021 | else None
1022 | )
1023 |
1024 | # 6.2 Optionally get Guidance Scale Embedding
1025 | timestep_cond = None
1026 | if self.unet.config.time_cond_proj_dim is not None:
1027 | guidance_scale_tensor = torch.tensor(
1028 | self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
1029 | timestep_cond = self.get_guidance_scale_embedding(
1030 | guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
1031 | ).to(device=device, dtype=latents.dtype)
1032 |
1033 | # 7. Denoising loop
1034 | num_warmup_steps = len(timesteps) - \
1035 | num_inference_steps * self.scheduler.order
1036 | self._num_timesteps = len(timesteps)
1037 | # intermediate_latents = []
1038 | with self.progress_bar(total=num_inference_steps) as progress_bar:
1039 | for i, t in enumerate(timesteps):
1040 | if self.interrupt:
1041 | continue
1042 |
1043 | # expand the latents if we are doing classifier free guidance
1044 | latent_model_input = torch.cat(
1045 | [latents] * 2) if self.do_classifier_free_guidance else latents
1046 | latent_model_input = self.scheduler.scale_model_input(
1047 | latent_model_input, t)
1048 |
1049 | # predict the noise residual
1050 | noise_pred = self.unet(
1051 | latent_model_input,
1052 | t,
1053 | encoder_hidden_states=prompt_embeds,
1054 | timestep_cond=timestep_cond,
1055 | cross_attention_kwargs=self.cross_attention_kwargs,
1056 | added_cond_kwargs=added_cond_kwargs,
1057 | return_dict=False,
1058 | )[0]
1059 |
1060 | # perform guidance
1061 | if self.do_classifier_free_guidance:
1062 | noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
1063 | noise_pred = noise_pred_uncond + self.guidance_scale * \
1064 | (noise_pred_text - noise_pred_uncond)
1065 |
1066 | if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
1067 | # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
1068 | noise_pred = rescale_noise_cfg(
1069 | noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
1070 |
1071 | # compute the previous noisy sample x_t -> x_t-1
1072 | scheduler_output = self.scheduler.step(
1073 | noise_pred, t, latents, **extra_step_kwargs, return_dict=True)
1074 | latents = scheduler_output.prev_sample
1075 | # if i % 10 == 0:
1076 | # intermediate_latents.append(scheduler_output.pred_original_sample)
1077 | if callback_on_step_end is not None:
1078 | callback_kwargs = {}
1079 | for k in callback_on_step_end_tensor_inputs:
1080 | callback_kwargs[k] = locals()[k]
1081 | callback_outputs = callback_on_step_end(
1082 | self, i, t, callback_kwargs)
1083 |
1084 | latents = callback_outputs.pop("latents", latents)
1085 | prompt_embeds = callback_outputs.pop(
1086 | "prompt_embeds", prompt_embeds)
1087 | negative_prompt_embeds = callback_outputs.pop(
1088 | "negative_prompt_embeds", negative_prompt_embeds)
1089 |
1090 | # call the callback, if provided
1091 | if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
1092 | progress_bar.update()
1093 | if callback is not None and i % callback_steps == 0:
1094 | step_idx = i // getattr(self.scheduler, "order", 1)
1095 | callback(step_idx, t, latents)
1096 |
1097 | if XLA_AVAILABLE:
1098 | xm.mark_step()
1099 |
1100 | image, dem = self.decode_rgbd(latents, generator, output_type)
1101 |
1102 | # intermediate = [self.decode_rgbd(latent,generator,output_type)for latent in intermediate_latents]
1103 |
1104 | # Offload all models
1105 | self.maybe_free_model_hooks()
1106 |
1107 | return image, dem
1108 |
--------------------------------------------------------------------------------
/models.py:
--------------------------------------------------------------------------------
1 | ###########################################################################
2 | # References:
3 | # https://github.com/huggingface/diffusers/
4 | ###########################################################################
5 |
6 | from dataclasses import dataclass
7 | from typing import Any, Dict, List, Optional, Tuple, Union
8 |
9 | import torch
10 | import torch.nn as nn
11 |
12 | from diffusers.configuration_utils import ConfigMixin, register_to_config
13 | from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
14 | from diffusers.loaders.single_file_model import FromOriginalModelMixin
15 | from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
16 | from diffusers.models.activations import get_activation
17 | from diffusers.models.attention_processor import (
18 | ADDED_KV_ATTENTION_PROCESSORS,
19 | CROSS_ATTENTION_PROCESSORS,
20 | Attention,
21 | AttentionProcessor,
22 | AttnAddedKVProcessor,
23 | AttnProcessor,
24 | FusedAttnProcessor2_0,
25 | )
26 | from diffusers.models.embeddings import (
27 | GaussianFourierProjection,
28 | GLIGENTextBoundingboxProjection,
29 | ImageHintTimeEmbedding,
30 | ImageProjection,
31 | ImageTimeEmbedding,
32 | TextImageProjection,
33 | TextImageTimeEmbedding,
34 | TextTimeEmbedding,
35 | TimestepEmbedding,
36 | Timesteps,
37 | )
38 | from diffusers.models.modeling_utils import ModelMixin
39 | from diffusers.models.unets.unet_2d_blocks import (
40 | get_down_block,
41 | get_mid_block,
42 | get_up_block,
43 | )
44 |
45 | logger = logging.get_logger(__name__) # pylint: disable=invalid-name
46 |
47 |
48 | @dataclass
49 | class UNet2DConditionOutput(BaseOutput):
50 | """
51 | The output of [`UNet2DConditionModel`].
52 |
53 | Args:
54 | sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`):
55 | The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
56 | """
57 |
58 | sample: torch.Tensor = None
59 |
60 |
61 | class UNetDEMConditionModel(
62 | ModelMixin, ConfigMixin, FromOriginalModelMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin
63 | ):
64 | r"""
65 | A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
66 | shaped output.
67 |
68 | This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
69 | for all models (such as downloading or saving).
70 |
71 | Parameters:
72 | sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
73 | Height and width of input/output sample.
74 | in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
75 | out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
76 | center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
77 | flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
78 | Whether to flip the sin to cos in the time embedding.
79 | freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
80 | down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
81 | The tuple of downsample blocks to use.
82 | mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
83 | Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
84 | `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
85 | up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
86 | The tuple of upsample blocks to use.
87 | only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
88 | Whether to include self-attention in the basic transformer blocks, see
89 | [`~models.attention.BasicTransformerBlock`].
90 | block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
91 | The tuple of output channels for each block.
92 | layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
93 | downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
94 | mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
95 | dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
96 | act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
97 | norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
98 | If `None`, normalization and activation layers is skipped in post-processing.
99 | norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
100 | cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
101 | The dimension of the cross attention features.
102 | transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
103 | The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
104 | [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
105 | [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
106 | reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
107 | The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
108 | blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
109 | [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
110 | [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
111 | encoder_hid_dim (`int`, *optional*, defaults to None):
112 | If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
113 | dimension to `cross_attention_dim`.
114 | encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
115 | If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
116 | embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
117 | attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
118 | num_attention_heads (`int`, *optional*):
119 | The number of attention heads. If not defined, defaults to `attention_head_dim`
120 | resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
121 | for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
122 | class_embed_type (`str`, *optional*, defaults to `None`):
123 | The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
124 | `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
125 | addition_embed_type (`str`, *optional*, defaults to `None`):
126 | Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
127 | "text". "text" will use the `TextTimeEmbedding` layer.
128 | addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
129 | Dimension for the timestep embeddings.
130 | num_class_embeds (`int`, *optional*, defaults to `None`):
131 | Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
132 | class conditioning with `class_embed_type` equal to `None`.
133 | time_embedding_type (`str`, *optional*, defaults to `positional`):
134 | The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
135 | time_embedding_dim (`int`, *optional*, defaults to `None`):
136 | An optional override for the dimension of the projected time embedding.
137 | time_embedding_act_fn (`str`, *optional*, defaults to `None`):
138 | Optional activation function to use only once on the time embeddings before they are passed to the rest of
139 | the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
140 | timestep_post_act (`str`, *optional*, defaults to `None`):
141 | The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
142 | time_cond_proj_dim (`int`, *optional*, defaults to `None`):
143 | The dimension of `cond_proj` layer in the timestep embedding.
144 | conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
145 | conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
146 | projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
147 | `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
148 | class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
149 | embeddings with the class embeddings.
150 | mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
151 | Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
152 | `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
153 | `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
154 | otherwise.
155 | """
156 |
157 | _supports_gradient_checkpointing = True
158 | _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]
159 |
160 | @register_to_config
161 | def __init__(
162 | self,
163 | sample_size: Optional[int] = None,
164 | in_channels: int = 8,
165 | out_channels: int = 8,
166 | center_input_sample: bool = False,
167 | flip_sin_to_cos: bool = True,
168 | freq_shift: int = 0,
169 | down_block_types: Tuple[str] = (
170 | "CrossAttnDownBlock2D",
171 | "CrossAttnDownBlock2D",
172 | "CrossAttnDownBlock2D",
173 | "DownBlock2D",
174 | ),
175 | mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
176 | up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
177 | only_cross_attention: Union[bool, Tuple[bool]] = False,
178 | block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
179 | layers_per_block: Union[int, Tuple[int]] = 2,
180 | downsample_padding: int = 1,
181 | mid_block_scale_factor: float = 1,
182 | dropout: float = 0.0,
183 | act_fn: str = "silu",
184 | norm_num_groups: Optional[int] = 32,
185 | norm_eps: float = 1e-5,
186 | cross_attention_dim: Union[int, Tuple[int]] = 1280,
187 | transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
188 | reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
189 | encoder_hid_dim: Optional[int] = None,
190 | encoder_hid_dim_type: Optional[str] = None,
191 | attention_head_dim: Union[int, Tuple[int]] = 8,
192 | num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
193 | dual_cross_attention: bool = False,
194 | use_linear_projection: bool = False,
195 | class_embed_type: Optional[str] = None,
196 | addition_embed_type: Optional[str] = None,
197 | addition_time_embed_dim: Optional[int] = None,
198 | num_class_embeds: Optional[int] = None,
199 | upcast_attention: bool = False,
200 | resnet_time_scale_shift: str = "default",
201 | resnet_skip_time_act: bool = False,
202 | resnet_out_scale_factor: float = 1.0,
203 | time_embedding_type: str = "positional",
204 | time_embedding_dim: Optional[int] = None,
205 | time_embedding_act_fn: Optional[str] = None,
206 | timestep_post_act: Optional[str] = None,
207 | time_cond_proj_dim: Optional[int] = None,
208 | conv_in_kernel: int = 3,
209 | conv_out_kernel: int = 3,
210 | projection_class_embeddings_input_dim: Optional[int] = None,
211 | attention_type: str = "default",
212 | class_embeddings_concat: bool = False,
213 | mid_block_only_cross_attention: Optional[bool] = None,
214 | cross_attention_norm: Optional[str] = None,
215 | addition_embed_type_num_heads: int = 64,
216 | ):
217 | super().__init__()
218 |
219 | self.sample_size = sample_size
220 |
221 | if num_attention_heads is not None:
222 | raise ValueError(
223 | "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
224 | )
225 |
226 | # If `num_attention_heads` is not defined (which is the case for most models)
227 | # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
228 | # The reason for this behavior is to correct for incorrectly named variables that were introduced
229 | # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
230 | # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
231 | # which is why we correct for the naming here.
232 | num_attention_heads = num_attention_heads or attention_head_dim
233 |
234 | # Check inputs
235 | self._check_config(
236 | down_block_types=down_block_types,
237 | up_block_types=up_block_types,
238 | only_cross_attention=only_cross_attention,
239 | block_out_channels=block_out_channels,
240 | layers_per_block=layers_per_block,
241 | cross_attention_dim=cross_attention_dim,
242 | transformer_layers_per_block=transformer_layers_per_block,
243 | reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
244 | attention_head_dim=attention_head_dim,
245 | num_attention_heads=num_attention_heads,
246 | )
247 |
248 | # input
249 | conv_in_padding = (conv_in_kernel - 1) // 2
250 | self.conv_in_img = nn.Conv2d(
251 | in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
252 | )
253 | self.conv_in_dem = nn.Conv2d(
254 | in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
255 | )
256 |
257 | # time
258 | time_embed_dim, timestep_input_dim = self._set_time_proj(
259 | time_embedding_type,
260 | block_out_channels=block_out_channels,
261 | flip_sin_to_cos=flip_sin_to_cos,
262 | freq_shift=freq_shift,
263 | time_embedding_dim=time_embedding_dim,
264 | )
265 |
266 | self.time_embedding = TimestepEmbedding(
267 | timestep_input_dim,
268 | time_embed_dim,
269 | act_fn=act_fn,
270 | post_act_fn=timestep_post_act,
271 | cond_proj_dim=time_cond_proj_dim,
272 | )
273 |
274 | self._set_encoder_hid_proj(
275 | encoder_hid_dim_type,
276 | cross_attention_dim=cross_attention_dim,
277 | encoder_hid_dim=encoder_hid_dim,
278 | )
279 |
280 | # class embedding
281 | self._set_class_embedding(
282 | class_embed_type,
283 | act_fn=act_fn,
284 | num_class_embeds=num_class_embeds,
285 | projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
286 | time_embed_dim=time_embed_dim,
287 | timestep_input_dim=timestep_input_dim,
288 | )
289 |
290 | self._set_add_embedding(
291 | addition_embed_type,
292 | addition_embed_type_num_heads=addition_embed_type_num_heads,
293 | addition_time_embed_dim=addition_time_embed_dim,
294 | cross_attention_dim=cross_attention_dim,
295 | encoder_hid_dim=encoder_hid_dim,
296 | flip_sin_to_cos=flip_sin_to_cos,
297 | freq_shift=freq_shift,
298 | projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
299 | time_embed_dim=time_embed_dim,
300 | )
301 |
302 | if time_embedding_act_fn is None:
303 | self.time_embed_act = None
304 | else:
305 | self.time_embed_act = get_activation(time_embedding_act_fn)
306 |
307 | self.down_blocks = nn.ModuleList([])
308 | self.up_blocks = nn.ModuleList([])
309 |
310 | if isinstance(only_cross_attention, bool):
311 | if mid_block_only_cross_attention is None:
312 | mid_block_only_cross_attention = only_cross_attention
313 |
314 | only_cross_attention = [only_cross_attention] * len(down_block_types)
315 |
316 | if mid_block_only_cross_attention is None:
317 | mid_block_only_cross_attention = False
318 |
319 | if isinstance(num_attention_heads, int):
320 | num_attention_heads = (num_attention_heads,) * len(down_block_types)
321 |
322 | if isinstance(attention_head_dim, int):
323 | attention_head_dim = (attention_head_dim,) * len(down_block_types)
324 |
325 | if isinstance(cross_attention_dim, int):
326 | cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
327 |
328 | if isinstance(layers_per_block, int):
329 | layers_per_block = [layers_per_block] * len(down_block_types)
330 |
331 | if isinstance(transformer_layers_per_block, int):
332 | transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
333 |
334 | if class_embeddings_concat:
335 | # The time embeddings are concatenated with the class embeddings. The dimension of the
336 | # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
337 | # regular time embeddings
338 | blocks_time_embed_dim = time_embed_dim * 2
339 | else:
340 | blocks_time_embed_dim = time_embed_dim
341 |
342 | # down
343 | output_channel = block_out_channels[0]
344 |
345 |
346 | for i, down_block_type in enumerate(down_block_types):
347 | input_channel = output_channel
348 | output_channel = block_out_channels[i]
349 | is_final_block = i == len(block_out_channels) - 1
350 | down_block_kwargs = {"down_block_type":down_block_type,
351 | "num_layers":layers_per_block[i],
352 | "transformer_layers_per_block":transformer_layers_per_block[i],
353 | "in_channels":input_channel,
354 | "out_channels":output_channel,
355 | "temb_channels":blocks_time_embed_dim,
356 | "add_downsample":not is_final_block,
357 | "resnet_eps":norm_eps,
358 | "resnet_act_fn":act_fn,
359 | "resnet_groups":norm_num_groups,
360 | "cross_attention_dim":cross_attention_dim[i],
361 | "num_attention_heads":num_attention_heads[i],
362 | "downsample_padding":downsample_padding,
363 | "dual_cross_attention":dual_cross_attention,
364 | "use_linear_projection":use_linear_projection,
365 | "only_cross_attention":only_cross_attention[i],
366 | "upcast_attention":upcast_attention,
367 | "resnet_time_scale_shift":resnet_time_scale_shift,
368 | "attention_type":attention_type,
369 | "resnet_skip_time_act":resnet_skip_time_act,
370 | "resnet_out_scale_factor":resnet_out_scale_factor,
371 | "cross_attention_norm":cross_attention_norm,
372 | "attention_head_dim":attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
373 | "dropout":dropout}
374 |
375 | if i == 0:
376 | self.head_img = get_down_block(**down_block_kwargs)
377 | # same architecture as the head_img but different weights
378 | self.head_dem = get_down_block(**down_block_kwargs)
379 | # elif i == 1:
380 | # down_block_kwargs["in_channels"] = input_channel *2 # concatenate the output of the head_img and head_dem
381 | # down_block = get_down_block(**down_block_kwargs)
382 | # self.down_blocks.append(down_block)
383 | else:
384 | down_block = get_down_block(**down_block_kwargs)
385 | self.down_blocks.append(down_block)
386 |
387 | # mid
388 | self.mid_block = get_mid_block(
389 | mid_block_type,
390 | temb_channels=blocks_time_embed_dim,
391 | in_channels=block_out_channels[-1],
392 | resnet_eps=norm_eps,
393 | resnet_act_fn=act_fn,
394 | resnet_groups=norm_num_groups,
395 | output_scale_factor=mid_block_scale_factor,
396 | transformer_layers_per_block=transformer_layers_per_block[-1],
397 | num_attention_heads=num_attention_heads[-1],
398 | cross_attention_dim=cross_attention_dim[-1],
399 | dual_cross_attention=dual_cross_attention,
400 | use_linear_projection=use_linear_projection,
401 | mid_block_only_cross_attention=mid_block_only_cross_attention,
402 | upcast_attention=upcast_attention,
403 | resnet_time_scale_shift=resnet_time_scale_shift,
404 | attention_type=attention_type,
405 | resnet_skip_time_act=resnet_skip_time_act,
406 | cross_attention_norm=cross_attention_norm,
407 | attention_head_dim=attention_head_dim[-1],
408 | dropout=dropout,
409 | )
410 |
411 | # count how many layers upsample the images
412 | self.num_upsamplers = 0
413 |
414 | # up
415 | reversed_block_out_channels = list(reversed(block_out_channels))
416 | reversed_num_attention_heads = list(reversed(num_attention_heads))
417 | reversed_layers_per_block = list(reversed(layers_per_block))
418 | reversed_cross_attention_dim = list(reversed(cross_attention_dim))
419 | reversed_transformer_layers_per_block = (
420 | list(reversed(transformer_layers_per_block))
421 | if reverse_transformer_layers_per_block is None
422 | else reverse_transformer_layers_per_block
423 | )
424 | only_cross_attention = list(reversed(only_cross_attention))
425 |
426 | output_channel = reversed_block_out_channels[0]
427 | for i, up_block_type in enumerate(up_block_types):
428 | is_final_block = i == len(block_out_channels) - 1
429 |
430 | prev_output_channel = output_channel
431 | output_channel = reversed_block_out_channels[i]
432 | input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
433 |
434 | # add upsample block for all BUT final layer
435 | if not is_final_block:
436 | add_upsample = True
437 | self.num_upsamplers += 1
438 | else:
439 | add_upsample = False
440 |
441 | up_block_kwargs = {"up_block_type":up_block_type,
442 | "num_layers":reversed_layers_per_block[i] + 1,
443 | "transformer_layers_per_block":reversed_transformer_layers_per_block[i],
444 | "in_channels":input_channel,
445 | "out_channels":output_channel,
446 | "prev_output_channel":prev_output_channel,
447 | "temb_channels":blocks_time_embed_dim,
448 | "add_upsample":add_upsample,
449 | "resnet_eps":norm_eps,
450 | "resnet_act_fn":act_fn,
451 | "resolution_idx":i,
452 | "resnet_groups":norm_num_groups,
453 | "cross_attention_dim":reversed_cross_attention_dim[i],
454 | "num_attention_heads":reversed_num_attention_heads[i],
455 | "dual_cross_attention":dual_cross_attention,
456 | "use_linear_projection":use_linear_projection,
457 | "only_cross_attention":only_cross_attention[i],
458 | "upcast_attention":upcast_attention,
459 | "resnet_time_scale_shift":resnet_time_scale_shift,
460 | "attention_type":attention_type,
461 | "resnet_skip_time_act":resnet_skip_time_act,
462 | "resnet_out_scale_factor":resnet_out_scale_factor,
463 | "cross_attention_norm":cross_attention_norm,
464 | "attention_head_dim":attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
465 | "dropout":dropout,}
466 |
467 | # if i == len(block_out_channels) - 2:
468 | # up_block_kwargs["in_channels"] = input_channel*2
469 | # up_block = get_up_block(**up_block_kwargs)
470 | # self.up_blocks.append(up_block)
471 |
472 | if is_final_block :
473 |
474 | self.head_out_img = get_up_block(**up_block_kwargs)
475 | self.head_out_dem = get_up_block(**up_block_kwargs)
476 |
477 | else :
478 | up_block = get_up_block(**up_block_kwargs)
479 | self.up_blocks.append(up_block)
480 |
481 |
482 | # out
483 | if norm_num_groups is not None:
484 | self.conv_norm_out_img = nn.GroupNorm(
485 | num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
486 | )
487 | self.conv_norm_out_dem = nn.GroupNorm(
488 | num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
489 | )
490 |
491 | self.conv_act = get_activation(act_fn)
492 |
493 | else:
494 | self.conv_norm_out_img = None
495 | self.conv_norm_out_dem = None
496 | self.conv_act = None
497 |
498 | conv_out_padding = (conv_out_kernel - 1) // 2
499 | self.conv_out_img = nn.Conv2d(
500 | block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
501 | )
502 | self.conv_out_dem = nn.Conv2d(
503 | block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
504 | )
505 |
506 | self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)
507 |
508 | def _check_config(
509 | self,
510 | down_block_types: Tuple[str],
511 | up_block_types: Tuple[str],
512 | only_cross_attention: Union[bool, Tuple[bool]],
513 | block_out_channels: Tuple[int],
514 | layers_per_block: Union[int, Tuple[int]],
515 | cross_attention_dim: Union[int, Tuple[int]],
516 | transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
517 | reverse_transformer_layers_per_block: bool,
518 | attention_head_dim: int,
519 | num_attention_heads: Optional[Union[int, Tuple[int]]],
520 | ):
521 | if len(down_block_types) != len(up_block_types):
522 | raise ValueError(
523 | f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
524 | )
525 |
526 | if len(block_out_channels) != len(down_block_types):
527 | raise ValueError(
528 | f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
529 | )
530 |
531 | if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
532 | raise ValueError(
533 | f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
534 | )
535 |
536 | if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
537 | raise ValueError(
538 | f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
539 | )
540 |
541 | if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
542 | raise ValueError(
543 | f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
544 | )
545 |
546 | if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
547 | raise ValueError(
548 | f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
549 | )
550 |
551 | if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
552 | raise ValueError(
553 | f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
554 | )
555 | if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
556 | for layer_number_per_block in transformer_layers_per_block:
557 | if isinstance(layer_number_per_block, list):
558 | raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
559 |
560 | def _set_time_proj(
561 | self,
562 | time_embedding_type: str,
563 | block_out_channels: int,
564 | flip_sin_to_cos: bool,
565 | freq_shift: float,
566 | time_embedding_dim: int,
567 | ) -> Tuple[int, int]:
568 | if time_embedding_type == "fourier":
569 | time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
570 | if time_embed_dim % 2 != 0:
571 | raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
572 | self.time_proj = GaussianFourierProjection(
573 | time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
574 | )
575 | timestep_input_dim = time_embed_dim
576 | elif time_embedding_type == "positional":
577 | time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
578 |
579 | self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
580 | timestep_input_dim = block_out_channels[0]
581 | else:
582 | raise ValueError(
583 | f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
584 | )
585 |
586 | return time_embed_dim, timestep_input_dim
587 |
588 | def _set_encoder_hid_proj(
589 | self,
590 | encoder_hid_dim_type: Optional[str],
591 | cross_attention_dim: Union[int, Tuple[int]],
592 | encoder_hid_dim: Optional[int],
593 | ):
594 | if encoder_hid_dim_type is None and encoder_hid_dim is not None:
595 | encoder_hid_dim_type = "text_proj"
596 | self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
597 | logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
598 |
599 | if encoder_hid_dim is None and encoder_hid_dim_type is not None:
600 | raise ValueError(
601 | f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
602 | )
603 |
604 | if encoder_hid_dim_type == "text_proj":
605 | self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
606 | elif encoder_hid_dim_type == "text_image_proj":
607 | # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
608 | # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
609 | # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
610 | self.encoder_hid_proj = TextImageProjection(
611 | text_embed_dim=encoder_hid_dim,
612 | image_embed_dim=cross_attention_dim,
613 | cross_attention_dim=cross_attention_dim,
614 | )
615 | elif encoder_hid_dim_type == "image_proj":
616 | # Kandinsky 2.2
617 | self.encoder_hid_proj = ImageProjection(
618 | image_embed_dim=encoder_hid_dim,
619 | cross_attention_dim=cross_attention_dim,
620 | )
621 | elif encoder_hid_dim_type is not None:
622 | raise ValueError(
623 | f"`encoder_hid_dim_type`: {encoder_hid_dim_type} must be None, 'text_proj', 'text_image_proj', or 'image_proj'."
624 | )
625 | else:
626 | self.encoder_hid_proj = None
627 |
628 | def _set_class_embedding(
629 | self,
630 | class_embed_type: Optional[str],
631 | act_fn: str,
632 | num_class_embeds: Optional[int],
633 | projection_class_embeddings_input_dim: Optional[int],
634 | time_embed_dim: int,
635 | timestep_input_dim: int,
636 | ):
637 | if class_embed_type is None and num_class_embeds is not None:
638 | self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
639 | elif class_embed_type == "timestep":
640 | self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
641 | elif class_embed_type == "identity":
642 | self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
643 | elif class_embed_type == "projection":
644 | if projection_class_embeddings_input_dim is None:
645 | raise ValueError(
646 | "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
647 | )
648 | # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
649 | # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
650 | # 2. it projects from an arbitrary input dimension.
651 | #
652 | # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
653 | # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
654 | # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
655 | self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
656 | elif class_embed_type == "simple_projection":
657 | if projection_class_embeddings_input_dim is None:
658 | raise ValueError(
659 | "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
660 | )
661 | self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
662 | else:
663 | self.class_embedding = None
664 |
665 | def _set_add_embedding(
666 | self,
667 | addition_embed_type: str,
668 | addition_embed_type_num_heads: int,
669 | addition_time_embed_dim: Optional[int],
670 | flip_sin_to_cos: bool,
671 | freq_shift: float,
672 | cross_attention_dim: Optional[int],
673 | encoder_hid_dim: Optional[int],
674 | projection_class_embeddings_input_dim: Optional[int],
675 | time_embed_dim: int,
676 | ):
677 | if addition_embed_type == "text":
678 | if encoder_hid_dim is not None:
679 | text_time_embedding_from_dim = encoder_hid_dim
680 | else:
681 | text_time_embedding_from_dim = cross_attention_dim
682 |
683 | self.add_embedding = TextTimeEmbedding(
684 | text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
685 | )
686 | elif addition_embed_type == "text_image":
687 | # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
688 | # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
689 | # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
690 | self.add_embedding = TextImageTimeEmbedding(
691 | text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
692 | )
693 | elif addition_embed_type == "text_time":
694 | self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
695 | self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
696 | elif addition_embed_type == "image":
697 | # Kandinsky 2.2
698 | self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
699 | elif addition_embed_type == "image_hint":
700 | # Kandinsky 2.2 ControlNet
701 | self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
702 | elif addition_embed_type is not None:
703 | raise ValueError(
704 | f"`addition_embed_type`: {addition_embed_type} must be None, 'text', 'text_image', 'text_time', 'image', or 'image_hint'."
705 | )
706 |
707 | def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
708 | if attention_type in ["gated", "gated-text-image"]:
709 | positive_len = 768
710 | if isinstance(cross_attention_dim, int):
711 | positive_len = cross_attention_dim
712 | elif isinstance(cross_attention_dim, (list, tuple)):
713 | positive_len = cross_attention_dim[0]
714 |
715 | feature_type = "text-only" if attention_type == "gated" else "text-image"
716 | self.position_net = GLIGENTextBoundingboxProjection(
717 | positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
718 | )
719 |
720 | @property
721 | def attn_processors(self) -> Dict[str, AttentionProcessor]:
722 | r"""
723 | Returns:
724 | `dict` of attention processors: A dictionary containing all attention processors used in the model with
725 | indexed by its weight name.
726 | """
727 | # set recursively
728 | processors = {}
729 |
730 | def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
731 | if hasattr(module, "get_processor"):
732 | processors[f"{name}.processor"] = module.get_processor()
733 |
734 | for sub_name, child in module.named_children():
735 | fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
736 |
737 | return processors
738 |
739 | for name, module in self.named_children():
740 | fn_recursive_add_processors(name, module, processors)
741 |
742 | return processors
743 |
744 | def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
745 | r"""
746 | Sets the attention processor to use to compute attention.
747 |
748 | Parameters:
749 | processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
750 | The instantiated processor class or a dictionary of processor classes that will be set as the processor
751 | for **all** `Attention` layers.
752 |
753 | If `processor` is a dict, the key needs to define the path to the corresponding cross attention
754 | processor. This is strongly recommended when setting trainable attention processors.
755 |
756 | """
757 | count = len(self.attn_processors.keys())
758 |
759 | if isinstance(processor, dict) and len(processor) != count:
760 | raise ValueError(
761 | f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
762 | f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
763 | )
764 |
765 | def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
766 | if hasattr(module, "set_processor"):
767 | if not isinstance(processor, dict):
768 | module.set_processor(processor)
769 | else:
770 | module.set_processor(processor.pop(f"{name}.processor"))
771 |
772 | for sub_name, child in module.named_children():
773 | fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
774 |
775 | for name, module in self.named_children():
776 | fn_recursive_attn_processor(name, module, processor)
777 |
778 | def set_default_attn_processor(self):
779 | """
780 | Disables custom attention processors and sets the default attention implementation.
781 | """
782 | if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
783 | processor = AttnAddedKVProcessor()
784 | elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
785 | processor = AttnProcessor()
786 | else:
787 | raise ValueError(
788 | f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
789 | )
790 |
791 | self.set_attn_processor(processor)
792 |
793 | def set_attention_slice(self, slice_size: Union[str, int, List[int]] = "auto"):
794 | r"""
795 | Enable sliced attention computation.
796 |
797 | When this option is enabled, the attention module splits the input tensor in slices to compute attention in
798 | several steps. This is useful for saving some memory in exchange for a small decrease in speed.
799 |
800 | Args:
801 | slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
802 | When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
803 | `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
804 | provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
805 | must be a multiple of `slice_size`.
806 | """
807 | sliceable_head_dims = []
808 |
809 | def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
810 | if hasattr(module, "set_attention_slice"):
811 | sliceable_head_dims.append(module.sliceable_head_dim)
812 |
813 | for child in module.children():
814 | fn_recursive_retrieve_sliceable_dims(child)
815 |
816 | # retrieve number of attention layers
817 | for module in self.children():
818 | fn_recursive_retrieve_sliceable_dims(module)
819 |
820 | num_sliceable_layers = len(sliceable_head_dims)
821 |
822 | if slice_size == "auto":
823 | # half the attention head size is usually a good trade-off between
824 | # speed and memory
825 | slice_size = [dim // 2 for dim in sliceable_head_dims]
826 | elif slice_size == "max":
827 | # make smallest slice possible
828 | slice_size = num_sliceable_layers * [1]
829 |
830 | slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
831 |
832 | if len(slice_size) != len(sliceable_head_dims):
833 | raise ValueError(
834 | f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
835 | f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
836 | )
837 |
838 | for i in range(len(slice_size)):
839 | size = slice_size[i]
840 | dim = sliceable_head_dims[i]
841 | if size is not None and size > dim:
842 | raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
843 |
844 | # Recursively walk through all the children.
845 | # Any children which exposes the set_attention_slice method
846 | # gets the message
847 | def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
848 | if hasattr(module, "set_attention_slice"):
849 | module.set_attention_slice(slice_size.pop())
850 |
851 | for child in module.children():
852 | fn_recursive_set_attention_slice(child, slice_size)
853 |
854 | reversed_slice_size = list(reversed(slice_size))
855 | for module in self.children():
856 | fn_recursive_set_attention_slice(module, reversed_slice_size)
857 |
858 | def _set_gradient_checkpointing(self, module, value=False):
859 | if hasattr(module, "gradient_checkpointing"):
860 | module.gradient_checkpointing = value
861 |
862 | def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
863 | r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
864 |
865 | The suffixes after the scaling factors represent the stage blocks where they are being applied.
866 |
867 | Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
868 | are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
869 |
870 | Args:
871 | s1 (`float`):
872 | Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
873 | mitigate the "oversmoothing effect" in the enhanced denoising process.
874 | s2 (`float`):
875 | Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
876 | mitigate the "oversmoothing effect" in the enhanced denoising process.
877 | b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
878 | b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
879 | """
880 | for i, upsample_block in enumerate(self.up_blocks):
881 | setattr(upsample_block, "s1", s1)
882 | setattr(upsample_block, "s2", s2)
883 | setattr(upsample_block, "b1", b1)
884 | setattr(upsample_block, "b2", b2)
885 |
886 | def disable_freeu(self):
887 | """Disables the FreeU mechanism."""
888 | freeu_keys = {"s1", "s2", "b1", "b2"}
889 | for i, upsample_block in enumerate(self.up_blocks):
890 | for k in freeu_keys:
891 | if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
892 | setattr(upsample_block, k, None)
893 |
894 | def fuse_qkv_projections(self):
895 | """
896 | Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
897 | are fused. For cross-attention modules, key and value projection matrices are fused.
898 |
899 |