| --- |
| license: apache-2.0 |
| base_model: |
| - Wan-AI/Wan2.1-T2V-1.3B |
| - Wan-AI/Wan2.1-T2V-14B |
| pipeline_tag: text-to-video |
| --- |
| # rCM: Score-Regularized Continuous-Time Consistency Model |
| # Causal-rCM: Teacher-Forcing meets Self-Forcing in Autoregressive Diffusion Distillation for Streaming Video Generation and Interactive World Models |
| [**Paper**](https://arxiv.org/abs/2510.08431) | [**Website**](https://research.nvidia.com/labs/dir/rcm) | [**Code**](https://github.com/NVlabs/rcm) |
|
|
| This repo holds converted Wan official checkpoints in (Causal-)rCM/TurboDiffusion style. |
|
|
| Specifically, (Causal-)rCM equivalently replaces the `Conv3d` layer in the original Wan with a `Linear` layer for patch embedding, facilitating further optimization. The layer weight is directly reshaped without value change, e.g., from shape [5120, 16, 1, 2, 2] (Conv3d) to shape [5120, 64] (Linear). |
|
|
| ## Citation |
|
|
| ``` |
| @article{zheng2025rcm, |
| title={Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistency}, |
| author={Zheng, Kaiwen and Wang, Yuji and Ma, Qianli and Chen, Huayu and Zhang, Jintao and Balaji, Yogesh and Chen, Jianfei and Liu, Ming-Yu and Zhu, Jun and Zhang, Qinsheng}, |
| journal={arXiv preprint arXiv:2510.08431}, |
| year={2025} |
| } |
| @article{zheng2026causal, |
| title={Causal-rCM: Teacher-Forcing meets Self-Forcing in Autoregressive Diffusion Distillation for Streaming Video Generation and Interactive World Models}, |
| author={Zheng, Kaiwen and He, Guande and Zhao, Min and Zhu, Hongzhou and Zhang, Jintao and Chen, Huayu and Chen, Jianfei and Lin, Chen-Hsuan and Liu, Ming-Yu and Zhu, Jun and Ma, Qianli}, |
| journal={TODO}, |
| year={2026} |
| } |
| ``` |
