---
license: apache-2.0
base_model:
- Wan-AI/Wan2.1-T2V-1.3B
- Wan-AI/Wan2.1-T2V-14B
tags:
- video-generation
- autoregressive
- text-to-video
---

# Clean Forcing — adapted base checkpoints

Checkpoints for **Clean Forcing: Drift-Resistant Autoregressive Video Diffusion with a Frozen Base**:
the causally adapted Wan2.1 bases (repo root) and the 14 drift-corrector LoRAs (`weights/`, rank 16 on
self-attention q/k/v/o; `weights/README.md` maps each LoRA to its base and paper row).

```bash
# from the code repo root: LoRAs land in weights/, bases in the repo root
huggingface-cli download illustro1/clean-forcing --local-dir .
mv adapted_base_4000.pt self_forcing/wan_cache/
```

| File | Base | Size | Used for |
|---|---|---|---|
| `adapted_base_4000.pt` | Wan2.1-T2V-1.3B + rank-64 causal adapter (merged, step 4K) | 0.57 GB | all 1.3B results (Table 1 and ablations) |
| `wan14b_adapted_base_4000.pt` | Wan2.1-T2V-14B + rank-64 causal adapter (merged, step 4K) | 8.4 GB | the 14B scale-transfer appendix |

Load the merged weights with `torch.load(path)["merged"]` into the block-causal pipeline (see the code repo's `RUN.md`).
Verify downloads with `sha256sum -c SHA256SUMS` (bases) and `cd weights && sha256sum -c SHA256SUMS` (LoRAs).

- Code: https://github.com/wnqw/clean_forcing
- Project page and paper: https://clean-forcing.github.io/

Both checkpoints are derived from Wan2.1 (Apache-2.0) and are released under the same license.