search

kohya-ss / sd-scripts

Apache License 2.0
5.19k stars 862 forks source link

[Feature Request] Implement LECO to train LoRA for erasing concepts from diffusion models #833

Open Poiuytrezay1 opened 1 year ago

Poiuytrezay1 commented 1 year ago

See https://github.com/p1atdev/LECO, based off https://github.com/rohitgandikota/erasing/tree/main.

This serves to train "sliders"-like LoRA, and I think it would be great addition to the sd-scripts repo.

kohya-ss commented 1 year ago

Thank you for suggestion. However I think the LECO repository is well implemented and documented, and easy to use. So I do not think LECO training needs to be implemented in this repository.

Poiuytrezay1 commented 1 year ago

Thank you for suggestion. However I think the LECO repository is well implemented and documented, and easy to use. So I do not think LECO training needs to be implemented in this repository.

I agree that LECO is well-implemented and documented. However, it's missing many of the features offered in the sd-scripts repository (e.g. gradient checkpointing, min SNR gamma, noise offset, clip skip...). Some of those are proposed by https://github.com/ostris/ai-toolkit, but it is not compatible with NAI checkpoints.

kohya-ss commented 1 year ago

Thank you for clarification. I agree that some of features in sd-scripts are also useful for LECO training. However, it will take a lot of work to implement LECO training. I would like to prioritize other tasks at this time. I would appreciate your understanding.