Wan Video Family: Open Source AI Video Generation by Alibaba Cloud
ComfyUI Wiki
Alibaba Cloud's open-source video generation series, from Wan2.1 to Wan2.2, with high-quality text-to-video, image-to-video and MoE architecture.
W
Wan by Alibaba Cloud
VideoText-to-VideoImage-to-VideoOpen SourceOpen-source video generation model series from Alibaba Cloud (Wan AI). From the initial Wan2.1 (Apache 2.0 licensed, 14B/1.3B parameters, T2V and I2V) to the next-generation Wan2.2 with MoE architecture (14B/5B parameters, hybrid T2V+I2V, film-level aesthetic control).
Versions
| Version | Architecture | Parameters | Highlights |
|---|---|---|---|
| Wan2.1 | Diffusion Transformer | 14B / 1.3B | Apache 2.0, T2V + I2V, 480P/720P, GGUF & Kijai support |
| Wan2.2 | MoE (Mixture of Experts) | 14B / 5B | Apache 2.0, hybrid TI2V, film-level aesthetic, first-last frame |
| Wan2.2 Dancer | MoE + Hierarchical | 14B | Apache 2.0, music-to-dance, minute-long 720p/30fps, 5 dance genres |
| Wan-Animate-2 | Dual-branch DiT | 14B | Apache 2.0, end-to-end character animation from driving video, text-driven viewpoint control, multi-character, real-time Lite |
Select a generation for tutorials and related news. Curated weight downloads live in the Model Hub catalog.
Ecosystem pages
Projects built on Wan. Each page covers what it does, ComfyUI support and where to get the weights.
| Page | What it does |
|---|---|
| AccVideo | Text-to-video on Wan2.1, up to 8x faster via synthetic-dataset distillation |
| Anisora | Anime-style video generation system from Bilibili |
| ATI | Trajectory control for object, region and camera motion |
| BindWeave | Subject-consistent video for single and multi-subject prompts |
| CausVid | Causal few-step video generation (CVPR 2025) |
| ChronoEdit | Image and video editing handled as video generation (NVIDIA) |
| EchoShot | Multi-shot portraits from several reference images |
| FantasyPortrait | Audio-driven portrait animation with several characters |
| FastVideo | Distillation LoRAs for real-time Wan2.1 and Wan2.2 5B |
| FlashVSR | Real-time video super-resolution at about 17 FPS |
| FLF2V | First and last frame to a looping 720p clip |
| InfiniteTalk | Talking-head avatars from a single image |
| Lightx2v | CFG and step distillation LoRAs, rank 4 to 256 |
| Lynx | IP-Adapter style subject-driven video (KwaiVGI) |
| MoviiGen | Cinematic film-style fine-tune, 14B in fp16 or fp8 |
| Phantom | Subject-consistent video in 14B and 1.3B (ByteDance) |
| SteadyDancer | Human dance animation that keeps the first frame |
| UniLumos | Relighting for both images and video (Alibaba DAMO) |
| Wan-Move | Point-level motion control via latent trajectories (NeurIPS 2025) |
Comments
Sign in with GitHub to join the discussion.