mirror of
https://github.com/huggingface/diffusers.git
synced 2026-01-27 17:22:53 +03:00
8e963d1c2a
* 7529 do not disable autocast for cuda devices * Remove typecasting error check for non-mps platforms, as a correct autocast implementation makes it a non-issue * add autocast fix to other training examples * disable native_amp for dreambooth (sdxl) * disable native_amp for pix2pix (sdxl) * remove tests from remaining files * disable native_amp on huggingface accelerator for every training example that uses it * convert more usages of autocast to nullcontext, make style fixes * make style fixes * style. * Empty-Commit --------- Co-authored-by: bghira <bghira@users.github.com> Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Diffusion Model Alignment Using Direct Preference Optimization
This directory provides LoRA implementations of Diffusion DPO proposed in DiffusionModel Alignment Using Direct Preference Optimization by Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik.
We provide implementations for both Stable Diffusion (SD) and Stable Diffusion XL (SDXL). The original checkpoints are available at the URLs below:
💡 Note: The scripts are highly experimental and were only tested on low-data regimes. Proceed with caution. Feel free to let us know about your findings via GitHub issues.
SD training command
accelerate launch train_diffusion_dpo.py \
--pretrained_model_name_or_path=runwayml/stable-diffusion-v1-5 \
--output_dir="diffusion-dpo" \
--mixed_precision="fp16" \
--dataset_name=kashif/pickascore \
--resolution=512 \
--train_batch_size=16 \
--gradient_accumulation_steps=2 \
--gradient_checkpointing \
--use_8bit_adam \
--rank=8 \
--learning_rate=1e-5 \
--report_to="wandb" \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=10000 \
--checkpointing_steps=2000 \
--run_validation --validation_steps=200 \
--seed="0" \
--report_to="wandb" \
--push_to_hub
SDXL training command
accelerate launch train_diffusion_dpo_sdxl.py \
--pretrained_model_name_or_path=stabilityai/stable-diffusion-xl-base-1.0 \
--pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
--output_dir="diffusion-sdxl-dpo" \
--mixed_precision="fp16" \
--dataset_name=kashif/pickascore \
--train_batch_size=8 \
--gradient_accumulation_steps=2 \
--gradient_checkpointing \
--use_8bit_adam \
--rank=8 \
--learning_rate=1e-5 \
--report_to="wandb" \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=2000 \
--checkpointing_steps=500 \
--run_validation --validation_steps=50 \
--seed="0" \
--report_to="wandb" \
--push_to_hub
SDXL Turbo training command
accelerate launch train_diffusion_dpo_sdxl.py \
--pretrained_model_name_or_path=stabilityai/sdxl-turbo \
--pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
--output_dir="diffusion-sdxl-turbo-dpo" \
--mixed_precision="fp16" \
--dataset_name=kashif/pickascore \
--train_batch_size=8 \
--gradient_accumulation_steps=2 \
--gradient_checkpointing \
--use_8bit_adam \
--rank=8 \
--learning_rate=1e-5 \
--report_to="wandb" \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=2000 \
--checkpointing_steps=500 \
--run_validation --validation_steps=50 \
--seed="0" \
--report_to="wandb" \
--is_turbo --resolution 512 \
--push_to_hub
Acknowledgements
This is based on the amazing work done by Bram here for Diffusion DPO: https://github.com/bram-w/trl/blob/dpo/.