Wan 2.2 in one paragraph: Wan is a family of open video models from Alibaba’s Wan team. Wan 2.2, released July 28, 2025, is free to download under the Apache 2.0 license, which allows commercial use, and runs on your own computer. The easiest way to use it is ComfyUI, which has official Wan 2.2 workflows. The small 5B model makes 720p video at 24 fps and, according to ComfyUI, fits on an 8 GB graphics card with offloading. The larger 14B models give better results but need far more memory.
Based on the official Wan and ComfyUI documentation, checked September 29, 2026. We haven’t benchmarked Wan on our own hardware yet; test results will be added to this guide.
Why Use Wan Instead of Kling or Veo?
- It’s free. No credits or subscriptions. You pay only for your own hardware or rented GPU time.
- It’s private. Everything runs on your machine; nothing is uploaded.
- No content queue or daily limits. Generate as much as your GPU can handle.
- You can customize it with LoRAs and ComfyUI nodes, like Stable Diffusion for images.
The trade-offs: setup takes an afternoon, generation is slower than cloud tools on most home GPUs, and Wan 2.2 doesn’t generate its own sound. The speech-to-video model takes audio you provide; it doesn’t create it. If you want dialogue and sound effects out of the box, see our Kling 3.0 or Veo 3 guides.
Which Wan 2.2 Model Fits Your GPU
| Model | What it does | Hardware (official) |
|---|---|---|
| TI2V-5B | Text-to-video and image-to-video in one model, 720p at 24 fps | 24 GB VRAM (such as an RTX 4090) in Wan’s own code; ComfyUI says it “should fit well on 8GB vram” with its offloading |
| T2V-A14B | Text-to-video, 480p and 720p | 80 GB on a single GPU in Wan’s own code |
| I2V-A14B | Image-to-video, 480p and 720p | 80 GB on a single GPU in Wan’s own code |
| S2V-14B | Speech-to-video: animates a character to audio you supply | 80 GB on a single GPU |
| Animate-14B | Takes a video and a character image, and either animates your character with the video’s motion or swaps them into the video | See the repository |
The 14B figures come from Wan’s own scripts. ComfyUI uses smaller fp8 files and offloads to system memory, so the 14B workflows run on less, but slowly. If you have a consumer GPU, start with TI2V-5B.
If you don’t have a strong GPU, rent one by the hour from a cloud GPU service, or use a hosted API. See fal.ai alternatives for artists for options.
How to Run Wan 2.2 in ComfyUI
1. Install or update ComfyUI
Wan 2.2 needs a recent ComfyUI. Install ComfyUI Desktop or update your existing install.
2. Open an official Wan 2.2 template
Open ComfyUI’s template browser and look in the video templates. There are four official Wan 2.2 templates:
- Wan 2.2 5B Video Generation (text or image to video)
- Wan 2.2 14B Text to Video
- Wan 2.2 14B Image to Video
- Wan 2.2 14B First-Last Frame to Video (animates between a start and end image)
3. Download the model files
ComfyUI’s tutorial lists the files for each template. For the 5B workflow:
| File | Folder |
|---|---|
wan2.2_ti2v_5B_fp16.safetensors | ComfyUI/models/diffusion_models/ |
wan2.2_vae.safetensors | ComfyUI/models/vae/ |
umt5_xxl_fp8_e4m3fn_scaled.safetensors | ComfyUI/models/text_encoders/ |
The 14B text-to-video workflow uses two diffusion files, wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors and wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors, plus wan_2.1_vae.safetensors and the same text encoder. The 14B models are a mixture-of-experts design: one expert handles the early, noisy steps (overall layout) and the other the later steps (detail), which is why there are two files. For the image-to-video and first-last-frame templates, download the files that ComfyUI’s tutorial lists for those templates.
4. Set size and length, then run
Set resolution and frame count in the latent node (EmptyHunyuanLatentVideo for text-to-video, Wan22ImageToVideoLatent for the 5B image-to-video path). Wan clips are short: Wan’s own scripts default to 81 frames, which is about 3 to 5 seconds depending on the frame rate. Start at a low resolution and short length, then scale up once you like the motion.
Type your prompt, add a start image if you’re using image-to-video, and click Run.
How to Write Wan Prompts
Wan responds well to the same structure as other video models: subject, action, setting, camera, lighting and style. The official example prompt is short and concrete:
Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage.To get more control, add camera and light:
Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage. Low-angle tracking shot circling the ring, a single hard spotlight from above, smoke in the air, cheering crowd blurred in the dark background. Cinematic, high contrast.Tips from the Wan team:
- Longer prompts help. Wan’s docs say extending prompts “can effectively enrich the details in the generated videos.” Their own scripts can do this automatically with a Qwen language model (locally or via Alibaba’s Dashscope API). In ComfyUI you can do the same thing by hand, or ask an assistant like Claude or ChatGPT to expand a short idea into a detailed paragraph.
- Wan understands Chinese and English. For the first-last-frame model, the Wan 2.1 docs note that Chinese prompts work best, so it can be worth translating your prompt.
- Keep one clear action per clip. With only a few seconds of footage, one movement reads better than a sequence.
- Use image-to-video for control. Start from a still you made in your image tool of choice, and let the prompt describe only the motion.
Wan 2.2 vs Wan 2.5 and Later
Alibaba also offers newer Wan models (such as Wan 2.5) as a hosted service through Alibaba Cloud. Those are closed and paid, unlike the open 2.2 weights. When people say “Wan” in the ComfyUI world, they almost always mean the free, open 2.1 and 2.2 releases. Be wary of sites offering downloads of Wan versions that don’t appear on the official Wan-Video GitHub.
Wan 2.2 Checklist
- Start with the 5B template on a consumer GPU
- Put each file in the folder ComfyUI’s tutorial lists
- Test short and small, then scale up
- Write a detailed prompt, or expand a short one with an assistant
- Use a start image to control the look
- Add sound separately in your editor
Next Steps
- Build detailed prompts with the Video Prompt Builder
- Compare Wan with cloud tools in Best AI Video Generators
- Get to know ComfyUI’s image side with our Stable Diffusion profile