GC-ST-DPO diagnostic checkpoints
Best-step UNet checkpoints for the GC-ST-DPO paper experiments (gradient-attributed spatio-temporal-token credit assignment for diffusion preference optimization). All models trained on Pick-a-Pic v2, seed=0, eff-batch 1024 x 500 steps, beta=2000.
Layout
| path | backbone | init | note |
|---|---|---|---|
sd15/gc_dpo_full/ |
SD-1.5 | JaydenLu666/SD1.5-SFT | main method, best step 300 |
sd15/mask_dpo/ |
SD-1.5 | SFT | spatial-mask-only ablation, step 400 |
sd15/diff_dpo/ |
SD-1.5 | SFT | Diffusion-DPO reproduction, step 400 |
sd15/gc_dpo_full_random_w/ |
SD-1.5 | SFT | falsification control (shuffled W), step 500 |
sdxl/gc_dpo_full/ |
SDXL | base (bf16) | main method SDXL, step 500 |
sdxl/mask_dpo/ |
SDXL | base (bf16) | ablation SDXL, step 100 |
sdxl/sft_winners/ |
SDXL | base (bf16) | SFT-on-winners baseline (Wallace --sft) |
Each directory contains a bare UNet2DConditionModel (config.json +
diffusion_pytorch_model.safetensors, fp16). Load by overriding the UNet of
the matching base pipeline (runwayml/stable-diffusion-v1-5 /
stabilityai/stable-diffusion-xl-base-1.0).
Best-step was selected on held-out Pick-a-Pic validation prompts (500) by equal-weight mean rank of PickScore + HPS-v2.
Code + eval pipelines: https://github.com/JuntaoXu/GCPO
- Downloads last month
- -