Delcher/fewsteprl-sd35

FewStepRL — SD3.5-Medium checkpoints

Released checkpoints of FewStepRL: Fine-tuning Few-step Diffusion Models with Reinforcement Learning (image experiments, 4-step SD3.5-Medium flow map).

Paper (arXiv:2611.08240) · Code · Project page · Video checkpoints (Wan2.1)

Contents

<trial>/
  config.yaml                          # the config the run saved
  checkpoint-<step>/generator_ema.pt   # EMA generator: Stage-3 LoRA + delta_embedder
aux_vel/{mix,ocr}/
  config.yaml
  aux_v_transformer/                   # auxiliary v-predictor (diffusers SD3Transformer2DModel)
SHA256SUMS

trial	trainer	reward	step
meanflownft_mix	meanflow_nft	mixed	400
dmdnft_mix	dmdnft_rl	mixed	400
dmdawm_mix	dmdawm_rl	mixed	400
fewstepgrpo_mix	meanflowgrpo_rl	mixed	300
meanflowawm_mix	meanflowawm_rl	mixed	400
meanflowopsd_mix	meanflowopsd_rl	mixed	400
reflmx_mix	meanflowrefl_rl	mixed	400
flowgrpo_mix / awm_mix / nft_mix	original objectives (baselines)	mixed	400
meanflownft_ocr	meanflow_nft	OCR	200
dmdnft_ocr	dmdnft_rl	OCR	200
dmdawm_ocr	dmdawm_rl	OCR	400
fewstepgrpo_ocr	meanflowgrpo_rl	OCR	200
meanflowawm_ocr	meanflowawm_rl	OCR	400

Scores and per-checkpoint notes: checkpoints/README.md.

Usage

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

git clone https://github.com/EDAPINENUT/fewsteprl-sd35 && cd fewsteprl-sd35
bash download_scripts/download_assets.sh          # SD3.5-Medium, AnyFlow Stage-1/2 LoRAs, reward models
hf download Delcher/fewsteprl-sd35 --local-dir checkpoints --include "*.pt" "*.safetensors"
cd checkpoints && shasum -a 256 -c SHA256SUMS && cd ..

generator_ema.pt is applied on top of SD3.5-Medium + the AnyFlow Stage-1 and Stage-2 LoRAs, never alone:

bash scripts/eval/sample.sh meanflownft_mix       # 4-step sampling with a released checkpoint
bash scripts/eval/eval_ckpt.sh meanflownft_mix 8  # sample + score

The aux_vel/{mix,ocr} v-predictors are only needed to train the dmd* trials.

License

The weights are derivatives of SD3.5-Medium and are subject to the Stability AI Community License.

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}
}