Wan-Video /Wan2.2

About Wan2.2

Wan2.2 is a family of open video models you can run yourself. It handles text-to-video, image-to-video, speech-to-video and character animation. The A14B single-GPU commands want 80GB of VRAM. Prompt extension also needs a paid Dashscope API key.

Best for: GPU owners who want open video models running locally instead of through an API.

• Runs a Mixture-of-Experts (MoE) design: expert models split the denoising steps
• Labels aesthetic data by lighting, composition, contrast and color tone
• Generates 720P at 24fps with TI2V-5B, which runs on a 4090
• Turns speech into video with S2V-14B, plus optional CosyVoice text-to-speech
• Animates or replaces a character with Animate-14B, copying movement and expression
• Runs across many GPUs with FSDP and DeepSpeed Ulysses; works in ComfyUI and Diffusers

1. 1
   Clone the repo and enter it: the command below then cd Wan2.2

   git clone https://github.com/Wan-Video/Wan2.2.git

   Copy
2. 2
   Install dependencies: the command below (needs torch >= 2.4.0). For CosyVoice speech, also run the next command

   pip install -r requirements.txt

   Copy

   pip install -r requirements_s2v.txt

   Copy
3. 3
   Download weights: the command below

   huggingface-cli download Wan-AI/Wan2.2-T2V-A14B --local-dir ./Wan2.2-T2V-A14B

   Copy
4. 4
   Run text-to-video on one GPU: the command below

   python generate.py --task t2v-A14B --size 1280*720 --ckpt_dir ./Wan2.2-T2V-A14B --offload_model True --convert_model_dtype --prompt "..."

   Copy
5. 5
   For 8 GPUs: the command below

   torchrun --nproc_per_node=8 generate.py --task t2v-A14B --size 1280*720 --ckpt_dir ./Wan2.2-T2V-A14B --dit_fsdp --t5_fsdp --ulysses_size 8 --prompt "..."

   Copy
6. 6
   Optional: add the command below (needs DASH_API_KEY) or 'local_qwen'

   --use_prompt_extend --prompt_extend_method 'dashscope'

   Copy