Advanced camera control and style adaptation
Wan 2.2 is built on a Mixture of Experts (MoE) diffusion architecture — approximately 10 billion parameters arranged as specialized experts for high-noise and low-noise diffusion stages, producing more efficient and higher-quality output than the 14B monolithic architecture of Wan 2.1. The Video Animation Control Engine (VACE) is the central differentiating feature: explicit camera trajectory inputs, subject locking (keeping a defined subject stationary relative to camera movement), and background stabilization for controlled scene composition. Combined with a few-shot LoRA pipeline that adapts the model to a custom visual style using only 10–20 reference images, Wan 2.2 is the most configurable model in the Wan family. Licensed under Apache 2.0, the underlying model is open for commercial use.Capabilities
VACE camera controls
Explicit camera trajectory inputs — pans, zooms, focus pulls, and custom paths — for precise cinematographic control over the generated video.
Subject locking
Keep a defined subject visually stable relative to camera movement — useful for product showcase, character focus, and controlled scene composition.
Background stabilization
Stabilize the background while the subject moves, or vice versa — independent control over foreground and background motion.
Few-shot LoRA style adaptation
Adapt the model to a custom visual style using just 10–20 reference images via a few-shot LoRA pipeline — style consistency across generations.
MoE architecture
~10B parameter Mixture of Experts model — more efficient than 14B monolithic architectures, with specialized processing for different noise levels.
Native 1080p at 30 FPS
Full HD output at 30 frames per second without upscaling.
Specifications
| Feature | Details |
|---|---|
| Developer | Alibaba (Wan Video) |
| Architecture | Mixture of Experts (MoE), ~10B parameters |
| Resolution | Native 1080p |
| Frame rate | 30 FPS |
| Duration | Up to 5 seconds (T2V); multiple for I2V |
| Camera control | VACE — pans, zooms, focus pulls, custom paths |
| Style adaptation | Few-shot LoRA (10–20 images) |
| Audio | No native audio |
| License | Apache 2.0 |
How to use
Configure camera controls
Use VACE parameters to set your camera trajectory — specify pans, zoom direction, focus pull behavior, and any subject locking requirements.
Write your prompt
Describe the scene with detail on subjects, setting, style, and motion. Wan 2.2’s semantic understanding handles complex compositional instructions.
Prompting tips
- Be explicit with camera trajectory — “Camera slowly pans right while maintaining focus on the stationary subject” is processed as a VACE instruction, not just a suggestion.
- Describe background vs. foreground separately — “Background blurs softly while the subject remains sharp and stationary in frame center” activates subject locking and background stabilization.
- Use style references for LoRA — For brand content requiring a specific visual style, the few-shot LoRA pipeline can adapt the model to your reference aesthetic.
Example prompts
A product bottle sits on a frosted glass surface. The camera orbits slowly around it 180 degrees, keeping the bottle perfectly centered. Studio lighting, clean white background. 5 seconds, 1080p.
A wide shot of a mountain valley. The camera slowly pushes forward through a field of wildflowers, foreground flowers in focus, mountains slightly blurred in background. Golden hour light. 5 seconds.
Compare models
| Model | Camera control | Style adapt | Audio | Architecture | Best for |
|---|---|---|---|---|---|
| Wan 2.2 | VACE explicit | LoRA (10–20 imgs) | No | MoE 10B | Camera-precise, brand style |
| Wan 2.5 | Prompt-based | No | Yes | — | Audio-visual sync |
| Wan 2.6 | Prompt-based | No | Yes | — | Character reference, audio |
| Kling 2.5 Pro | Prompt-based | No | No | — | Fast, affordable 1080p |