---
language:
  - en
pipeline_tag: text-to-video
tags:
  - video-generation
  - autoregressive
  - wan
  - rmd
base_model:
  - Wan-AI/Wan2.1-T2V-1.3B
  - Wan-AI/Wan2.1-T2V-14B
---

# 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

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

Inference also needs the [Wan 1.3B assets](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B).
These checkpoints use the RMD code, not a standalone Diffusers pipeline.

See the [quick start](https://github.com/cjeen/RMD#quick-start) for generation
and the [training guide](https://github.com/cjeen/RMD#train-rmd)
for training. Causal ODE initialization is downloaded separately from
[Causal Forcing](https://huggingface.co/zhuhz22/Causal-Forcing).

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