RMD: Rollout-Marginal Distillation

Train on 5 seconds. Generate far beyond.

Pretrained weights for long-horizon autoregressive video generation with a four-step causal Wan 1.3B model.

Weights

File	Purpose
stage2/model.pt	Final model for inference (5.68 GB)
stage1/model.pt	Stage 1 model / initialization for stage 2
training/fake_score.pt	Adapted fake-score initialization
training/teacher_lora/adapter_model.bin	Wan 14B teacher LoRA

Download

hf download cjeen/RMD stage2/model.pt --local-dir checkpoints/rmd

Inference also needs the Wan 1.3B assets. These checkpoints use the RMD code, not a standalone Diffusers pipeline.

See the quick start for generation and the training guide for training. Causal ODE initialization is downloaded separately from Causal Forcing.

Upstream model and initialization licenses still apply; CODE_LICENSE covers the source code only. File checksums are listed in manifest.json.