Wan 2.2

Most open video models force a choice: quality or accessibility. Wan 2.2 refuses that trade-off. Alibaba's Tongyi Lab built a mixture-of-experts architecture that delivers cinematic motion and film-grade aesthetics while running on hardware you might already own. The weights are open, the license is Apache 2.0, and the results compete with closed systems costing far more to run. Released in July 2025, it quickly became a reference point for open video generation, with the weights freely available for research and commercial projects.

What Is Wan 2.2?

Wan 2.2 is an open-weights family of video generation models from Alibaba. Three variants cover the main workflows: T2V-A14B turns text prompts into video, I2V-A14B animates still images, and TI2V-5B handles both from a compact 5-billion-parameter package that fits on consumer GPUs like the RTX 4090.

The headline innovation is its MoE flow-matching design. Separate expert models handle high-noise and low-noise denoising stages, which means cleaner motion and sharper detail without ballooning compute costs. Training data grew substantially over Wan 2.1 — 65% more images, 83% more videos — and it shows in how the model handles complex scenes, multiple subjects, and stylized requests.

Key Features

1. Mixture-of-Experts Design

Specialized expert networks split the denoising workload, producing cleaner output at the same compute budget as conventional models.

2. True Open Weights

Download and self-host under Apache 2.0. Run locally, fine-tune, or integrate into your own pipeline with no vendor lock-in.

3. Cinematic Camera Control

Direct lighting, color, and framing with VACE 2.0 trajectory controls — pans, zooms, and focus pulls that feel intentionally directed.

4. Few-Shot Style Adaptation

Teach the model a custom look with just 10–20 reference images through an improved LoRA pipeline with visual blending sliders.

5. Bilingual Prompting

Write prompts in English or Chinese and get accurate in-frame text rendering in both languages.

6. Volumetric Effects

Fire, smoke, and dynamic lighting render with physical plausibility instead of the smeared approximations cheaper models produce.

7. Consumer GPU Friendly

The 5B variant generates 720p video at 24fps on a single high-end consumer card — no data center required.

Frequently Asked Questions

Can I use Wan 2.2 commercially?

Yes. The models are released under Apache 2.0, which permits commercial use, modification, and redistribution.

What hardware do I need to run it?

The TI2V-5B model runs on consumer GPUs with sufficient VRAM (a 4090-class card works); the larger A14B variants need more memory.

How long are the generated clips?

Up to 5 seconds per generation at 720p resolution and 24 frames per second.

How is it different from Wan 2.1?

The MoE architecture, native higher resolution, VACE 2.0 camera controls, and substantially larger training data are the major upgrades.

Does it work with ComfyUI?

Yes. The open-source community has built native ComfyUI support, and the models are available on Hugging Face and ModelScope.

What is Wan2.2-S2V?

A sibling model in the Wan 2.2 family for speech-to-video — it animates a portrait photo using an audio clip to create talking digital humans.

Conclusion

Wan 2.2 matters because it proves open models can lead. The MoE architecture delivers genuinely cinematic results, the license invites real commercial use, and the 5B variant puts that power on a desk instead of in a data center. Five-second clips won't replace a film crew, but for rapid visualization, short-form content, and research, this is the open video model to beat.