(Japanese) https://huggingface.co/piyohogeo/badanatomy-inhibitor/blob/main/README.jp.md
Abstract
This LoRA model is designed to suppress artistic anatomical errors in images generated by StableDiffusion. It is based on the CounterFeit-V2.5 model (https://huggingface.co/gsdf/Counterfeit-V2.5) and is trained with generated training data and the model. No instance tokens are required for usage. An alpha value of 1.0 for LoRA is too high for side effects, so it may be better to start with 0.5 to 0.75.
Example
Prompt: (best quality), super fine illustration, ((ultra-detailed)), contour deepening, Depth of field, 1girl, leg, <lora:badanatomy-inhibitor-v0.21312.3123.1:0.0>
Negative prompt: EasyNegative,(worst quality:1.5), (low quality:1.5)
Seed: 2430884677, 2795285915, 1763430179, 3990866940, 1595082783, 2178584239, 1670134450, 2627195931
Prompt: (best quality), super fine illustration, ((ultra-detailed)), contour deepening, Depth of field, 1girl, close up of hand, <lora:badanatomy-inhibitor-v0.21312.3123.1:0.0>
Negative prompt: EasyNegative, (worst quality:1.5), (low quality:1.5)
Seed: 2503513855, 763836568, 2225222013, 3669095192, 2727945938, 69933301, 4079409201, 955936219, 827443220, 638188775, 129799474
Prompt: (best quality), super fine illustration, ((ultra-detailed)), contour deepening, Depth of field, 1girl, close up of hands, <lora:badanatomy-inhibitor-v0.21312.3123.1:0.0>
Negative prompt: EasyNegative, bad-hands-5, (worst quality:1.5), (low quality:1.5)
Seed: 1052436550, 870994726, 1761127489, 1646083832, 2707827417, 986504348, 994879696, 773802065
Training Method
The training data is created by randomly generating a set of images using StableDiffusion, and then human selects the OK/NG (Good/Bad) images. During training, the OK images use the Mean Squared Error (MSE) loss, while the NG images use the MSE loss raised to the power of -α (where α is a constant > 0).
The source code used for the training is as follows (will not work alone due to dependencies) https://github.com/piyohogeo/sd_dreambooth_extension
The config used to train badanatomy-inhibitor-v0.21312.3123.1 is as follows db_config