Views
No views yet
pip install -r requirements.txtaccelerate.bash scripts/train/sd15_diffusion_dpo.sh--train_method selects Diffusion-DPO as the baseline. Choices : [diffusion-dpo, dspo, dmpo]--beta_dpo controls the DPO temperature or strength.--use_winner_preserving enables our SDPO safeguard that rescales only the loser branch’s backward signal to avoid increasing the winner loss to first order.--winner_preserving_mu sets the safeguard strength. Larger values are more conservative.--mixed_precision bf16 and --allow_tf32 improve throughput on recent NVIDIA GPUs.unet.pth checkpoint and will:prompts/:papv2.json is deduplicated from the Pick-a-Pic v2 test set to ensure prompts are uniquehpsv2.json and partiprompts.json are standard prompt suites used for qualitative and quantitative checks integrated from HPDv2 and Parti.bash scripts/eval/test_sd15.sh /path/to/your/unet.pthbash scripts/eval/test_sdxl.sh /path/to/your/unet.pth1# A/B win-rate comparison across all prompts from one group (papv2, hpsv2, partiprompts)
2# A.json / B.json are the generation manifests produced by your eval runs.
3
4bash scripts/eval/test_vs.sh \
5 --json_a path/to/A.json \
6 --json_b path/to/B.json \
7 --label_a "your label A" \
8 --label_b "your label B"1bash scripts/eval/test_vs.sh \
2 --json_a /path/to/sdxl/diffusion-dpo/hpsv2_seed0_1024x1024_50s_7.5cfg.json \
3 --json_b /path/to/sdxl/dmpo/hpsv2_seed0_1024x1024_50s_7.5cfg.json \
4 --label_a "diffusion_dpo_sdxl_hpsv2" \
5 --label_b "dmpo_sdxl_hpsv2"1@article{fu2025diffusion,
2 title={{Diffusion-SDPO}: Safeguarded Direct Preference Optimization for Diffusion Models},
3 author={Fu, Minghao and Wang, Guo-Hua and Cui, Tianyu and Chen, Qing-Guo and Xu, Zhao and Luo, Weihua and Zhang, Kaifu},
4 journal={arXiv:2511.03317},
5 year={2025}
6}