What Is Wan 2.2 and Should You Use It for AI Video?

Most AI video models can show you a cat on a skateboard. Wan 2.2 lets you decide how it’s lit, how the camera moves, and what kind of lens you’re pretending to shoot it with. In this post, we’ll break down what it does differently… and whether it’s actually worth your time.

All You Need to Know About Wan 2.2

Wan 2.2 is a Text to Video and Image to Video model released by Alibaba’s Tongyi Lab in July 2025. It builds on the foundation laid by Wan 2.1 but pushes things further, especially when it comes to motion quality, prompt responsiveness, and overall visual richness.

The biggest leap? Training data. Wan 2.2 wasn’t just fed more clips, it was fed better ones. The dataset grew by over 65% in still images and more than 80% in video, but it’s the labeling that matters. Each clip came annotated with precise visual metadata: lighting conditions, color grading, camera movement, framing, texture, and more.

This is why Wan 2.2 doesn’t just generate a man walking through a city at night. Because it knows the difference, it gives you a slow pan across wet pavement, moody sidelighting, and a soft lens flare from passing headlights.

A two-expert system

Under the hood, Wan 2.2 introduces a more flexible way of thinking. Instead of relying on one massive model to do everything, it uses a Mixture of Experts (MoE) system: two specialized components take turns handling different parts of the generation process.

The first expert sets the stage, laying down structure, movement, spatial depth. The second swoops in to refine the details, sharpen textures, and clean up the noise. It’s like having a cinematographer work alongside a compositor, with each playing to their strengths.

That split also makes the model more efficient. While the total architecture clocks in at 27 billion parameters, only a portion is active at any given time. So you get a large-model feel, without the GPU meltdown.

Camera-aware prompting

Wan 2.2 isn’t just built to respond to prompts, it’s built to take direction. Thanks to a built-in system called VACE 2.0 (short for Video Animation Control Engine), it understands and responds to motion language in a way most open models can’t.

That means prompts like:

• “Aerial orbit around a snowy cabin at dusk”
• “Handheld tracking shot through a neon-lit alley”
• “Zoom-in on a character’s face, warm backlight, shallow focus”

…actually work, more often than not. You can prompt for lighting style, camera type, lens effect, even mood, and get results that reflect that input. It’s not perfect (nothing in this space is), but it’s impressively good at hitting the tone you ask for.

Can you run it yourself?

Technically, yes. But unless you like debugging your own driver errors, you probably shouldn’t.

Wan 2.2 includes multiple variants, including a compact 5B hybrid version (TI2V‑5B) that can run on a single RTX 4090. But even then, you’ll need to manage compression settings, memory limits, model configs, and enough Python tools to fill a weekend.

So yes, it’s open source. But that doesn’t make it user-friendly.

Bottom line

Wan 2.2 doesn’t try to be everything, but what it does, it does surprisingly well, especially if you’re looking for more control over how your video looks, not just what it shows.