Delcher/fewsteprl-wan21

FewStepRL — Wan2.1-T2V-1.3B checkpoints

Released checkpoints of FewStepRL: Fine-tuning Few-step Diffusion Models with Reinforcement Learning (video experiments, 4-step AnyFlow-Wan2.1-T2V-1.3B flow map, 480×832, 81 frames).

Paper (arXiv:2611.08240) · Code · Project page · Image checkpoints (SD3.5)

Contents

<trial>/
  config.yaml                         # the config the run saved
  checkpoint-250/generator_ema.pt     # EMA generator: LoRA A/B + condition_embedder.delta_embedder
SHA256SUMS

trial	trainer	step
meanflownft	meanflow_nft	250
dmdnft	dmdnft_rl	250
dmdawm	dmdawm_rl	250
fewstepgrpo	meanflowgrpo_rl	250
meanflowawm	meanflowawm_rl	250
meanflowopsd	meanflowopsd_rl	250
reflmx	meanflowrefl_rl	250

All trials share the reward HPSv3 general + HPSv3 percentile + VideoAlign MQ + VideoAlign TA. Scores (training reward and VBench): checkpoints/README.md.

Usage

Download the weights into the checkpoints/ directory of the code repository:

git clone https://github.com/EDAPINENUT/fewsteprl-wan21 && cd fewsteprl-wan21
bash download_scripts/download_assets.sh          # AnyFlow-Wan, reward models, VBench weights
hf download Delcher/fewsteprl-wan21 --local-dir checkpoints --include "*.pt"
cd checkpoints && shasum -a 256 -c SHA256SUMS && cd ..

generator_ema.pt is applied on top of the full AnyFlow-Wan2.1-T2V-1.3B pipeline:

python inference.py scripts/eval/configs/sample_wan.yaml --prompt "..." \
  --override seed=0 mode=meanflow_nft meanflow_nft_path=checkpoints/dmdnft/checkpoint-250/generator_ema.pt
bash scripts/eval/eval_ckpt.sh dmdnft 8           # training rewards on the 256 test prompts

License

The weights are derivatives of nvidia/AnyFlow-Wan2.1-T2V-1.3B-Diffusers, whose license restricts the model and derivative weights to non-commercial use.

Citation

@article{lin2026fewsteprl,
  title   = {FewStepRL: Fine-tuning Few-step Diffusion Models with Reinforcement Learning},
  author  = {Lin, Haitao and Ren, Minsi and Li, Yang and Ge, Xingtong and
             Hu, Zheyuan and Hu, Peiyan and Wu, Tailin},
  journal = {arXiv preprint arXiv:2611.08240},
  year    = {2026}
}