├── 1.webp ├── 2.webp ├── 3.webp ├── 1xl.webp ├── 2xl.webp └── readme.md /1.webp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RimoChan/rimochan-LORA/HEAD/1.webp -------------------------------------------------------------------------------- /2.webp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RimoChan/rimochan-LORA/HEAD/2.webp -------------------------------------------------------------------------------- /3.webp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RimoChan/rimochan-LORA/HEAD/3.webp -------------------------------------------------------------------------------- /1xl.webp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RimoChan/rimochan-LORA/HEAD/1xl.webp -------------------------------------------------------------------------------- /2xl.webp: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/RimoChan/rimochan-LORA/HEAD/2xl.webp -------------------------------------------------------------------------------- /readme.md: -------------------------------------------------------------------------------- 1 | # 【rimochan-LORA】量产型电子莉沫酱！ 2 | 3 | 4 | 大家有用过Stable Diffusion吗？Stable Diffusion是1个生成图像的模型，嗯……感觉介绍很麻烦所以就不说了！ 5 | 6 | 总之我训练了1个LORA，可以用在Stable Diffusion上，来生成量产型电子莉沫酱图像！ 7 | 8 | 9 | ## 下载地址 10 | 11 | 因为GitHub不让我放超过100MB的文件，所以我把它们放到[release](https://github.com/RimoChan/rimochan-LORA/releases/download/%E5%A5%BD/rimo-lora-pack.zip)里啦。 12 | 13 | 14 | ## txt2img样例 15 | 16 | ![1.webp](1.webp) 17 | 18 | 19 | 看起来很像莉沫酱！ 20 | 21 | 但其实是因为模型是挑过的，虽然训练的时候是在novel_ai的模型上训的，但是测试下来发现Counterfeit-V3.0的效果最好，就用它的结果来做样例了。 22 | 23 | 使用的Prompt为: 24 | 25 | ``` 26 | rimochan, 27 | blush, 什么姿势, 室内还是室外, 28 | japanese clothes, kimono, print kimono, yellow kimono, yellow sleeves, strapless dress, dress, choker, obi, sash, detached sleeves 29 | ``` 30 | 31 | 32 | 完整的参数为: 33 | 34 | ```txt 35 | rimochan, 36 | blush, sitting, outdoors, 37 | japanese clothes, kimono, print kimono, yellow kimono, yellow sleeves, strapless dress, dress, choker, obi, sash, detached sleeves, 38 | 39 | Negative prompt: (worst quality, low quality:1.4), (realistic, lip, nose, tooth, rouge, lipstick, eyeshadow:1.0), (abs, muscular, rib:1.0), (blurry, greyscale, monochrome:1.0), text, title, logo, signature 40 | Steps: 50, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 512x704, Model hash: cbfba64e66, Model: Counterfeit-V3.0_fp16, VAE hash: ab33ad55c6, VAE: Counterfeit-V2.5.vae.pt, Lora hashes: "rimochan_v2: 135809c2d542", Eta: 0.68, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "sitting, standing, kneeling", Y Type: Prompt S/R, Y Values: "outdoors, indoors", Version: v1.6.0 41 | ``` 42 | 43 | 44 | ## 换衣服样例 45 | 46 | 我也测试了1下模型的泛化性，给莉沫酱换上各种衣服试试吧！ 47 | 48 | ![2.webp](2.webp) 49 | 50 | 51 | 看起来都很可爱，好耶！ 52 | 53 | 使用的Prompt为: 54 | 55 | ``` 56 | rimochan, 穿什么衣服, indoors 57 | ``` 58 | 59 | 完整的参数为: 60 | 61 | ``` 62 | rimochan, chef, indoors, 63 | 64 | Negative prompt: (worst quality, low quality:1.4), (realistic, lip, nose, tooth, rouge, lipstick, eyeshadow:1.0), (abs, muscular, rib:1.0), (blurry, greyscale, monochrome:1.0), text, title, logo, signature 65 | Steps: 50, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1, Size: 512x704, Model hash: cbfba64e66, Model: Counterfeit-V3.0_fp16, VAE hash: ab33ad55c6, VAE: Counterfeit-V2.5.vae.pt, Lora hashes: "rimochan_v2: 135809c2d542", Eta: 0.68, Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "chef,scientist,witch,assassin,priest,maid,magical girl,ninja", Version: v1.6.0 66 | ``` 67 | 68 | ## 关于模型版本和Prompt 69 | 70 | v1是比较早训的，原始训练集只有5张，做了1些裁剪放大和翻转之后增加到20张，效果也不错，不过有一些奇怪的问题。 71 | 72 | v2版本的模型加了几张新莉沫酱图，以及我弄了一些画师酱的杂图掺在数据集里做扩充。扩充部分的Prompt不带有`rimochan`，因此生成时Prompt需要加上`rimochan`才能稳定生成莉沫酱图片。 73 | 74 | 这个Prompt不是只是给人添麻烦的，它有用的地方是可以降低lora的刻板印象。像是v1版本的lora不能换发型，即使指定了`twintails`，模型还是会生成`low ponytail`的莉沫酱，而v2只要降低Prompt中`rimochan`的权重就可以让模型顺利生成`twintails`了，就像这样—— 75 | 76 | ![3.webp](3.webp) 77 | 78 | 降权后眼睛的颜色好像不是很对，没事，遇到这种情况就手动调整一下吧！ 79 | 80 | 81 | 使用的Prompt为: 82 | 83 | ``` 84 | (rimochan:权重), twintails, chef 85 | ``` 86 | 87 | 完整的参数为: 88 | 89 | ``` 90 | (rimochan:0.4), twintails, chef 91 | 92 | Negative prompt: (worst quality, low quality:1.4), (realistic, lip, nose, tooth, rouge, lipstick, eyeshadow:1.0), (abs, muscular, rib:1.0), (blurry, greyscale, monochrome:1.0), text, title, logo, signature 93 | Steps: 50, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1, Size: 448x704, Model hash: cbfba64e66, Model: Counterfeit-V3.0_fp16, VAE hash: ab33ad55c6, VAE: Counterfeit-V2.5.vae.pt, Lora hashes: "rimochan_v2: 135809c2d542", Script: X/Y/Z plot, X Type: Prompt S/R, X Values: "0.4,0.6,0.8,1.0", Version: v1.6.0 94 | ``` 95 | 96 | ## XL-LORA 97 | 98 | 因为最近我在用XL模型，所以还训练了1个XL模型用的LORA。 99 | 100 | 看起来也不错，样例图片是这样—— 101 | 102 | 换姿势: 103 | 104 | ![1xl.webp](1xl.webp) 105 | 106 | 换衣服: 107 | 108 | ![2xl.webp](2xl.webp) 109 | 110 | ## 结束 111 | 112 | 就这样，我要去当人工智能科学家了，大家88！ 113 | --------------------------------------------------------------------------------