CausVid: From Slow Bidirectional to Fast Causal Video Generation
CausVid by lightx2v/tianweiy adapts bidirectional diffusion transformers to causal autoregressive video generation on Wan2.1. CVPR 2025. Available in 14B fp8 with rank32 LoRAs.
CausVid
Wan2.1Video GenerationCausal GenerationAutoregressiveCVPR 2025Causal video generation model by lightx2v/tianweiy. Adapts pretrained bidirectional Wan2.1 diffusion transformers to fast autoregressive causal generation. CVPR 2025. Available in 14B fp8 with rank32 LoRAs.
| Developer | lightx2v / tianweiy |
| Architecture | Wan2.1-based with causal autoregressive attention |
| Model Sizes | 14B (fp8), plus LoRAs (rank32) |
| Capabilities | Real-time causal video generation, interactive applications |
Overview
CausVid (CVPR 2025) addresses a fundamental limitation of current video diffusion models: the bidirectional attention dependencies that prevent real-time interactive use. By adapting a pretrained Wan2.1 bidirectional diffusion transformer into a causal autoregressive transformer, CausVid enables frame-by-frame generation where each frame only depends on past frames rather than the entire sequence.
This approach dramatically reduces latency for interactive applications, making it suitable for real-time video generation scenarios while maintaining high output quality.
ComfyUI Integration
CausVid for Wan2.1 is available in fp8 precision along with rank32 LoRA adapters through the Kijai WanVideoWrapper for ComfyUI.
Available Variants
| Variant | Precision | Description |
|---|---|---|
| 14B Base | fp8 | Core causal generation model |
| LoRAs | rank32 | Fine-tuning adapters for causal generation |
Resources
| Resource | Link |
|---|---|
| GitHub Repository | github.com/tianweiy/CausVid |
| Project Page | causvid.github.io |
| Paper | arXiv:2412.07772 |
| Hugging Face (Original) | lightx2v/Wan2.1-T2V-14B-CausVid |
| Hugging Face (Kijai) | Kijai/WanVideo_comfy |
Guides and workflows related to this model series.