Wan 2.2

Highlights

• MoE architecture: two experts divide denoising stages, with clear quality improvement
• TI2V-5B uses a 16×16×4 high-compression VAE, 720p and VRAM-friendly
• Supports 5-second 480P/720P, with image-to-video/text-to-video/first-last-frame/S2V audio-driven all available
• Apache-2.0, the most worry-free tier for commercial use

Scope

• Good for: general image-to-video, talking-head/product shots
• Good for: projects requiring clean Apache-2.0 commercial use
• Not for: requiring long duration and strong narrative coherence

Tips

• LoRA for 5B and 14B are not interchangeable
• S2V is good for “lip-sync to audio”; do not use it for pure text-to-video
• When MoE switches experts, VRAM peak is higher; set batch to 1

Cautions

• MoE version's VRAM peak is more alarming than the nominal figure

Parameters

Duration

5 seconds

Resolution

480P / 720P

Mode

T2V, I2V, TI2V-5B, S2V (audio-driven)

License

Apache-2.0

Deployment

16GB conclusion

The 5B version is most suitable; A14B requires fp8/GGUF and repeatedly loads when switching experts.

VRAM note

A14B's MoE needs two experts rotating; on 16GB this causes frequent swap-in/swap-out, and actual speed is slower than the specs suggest.

Suggested starting point

First run the video_wan2_2_5B_ti2v template at 720p/5 seconds; this is the sweet spot for 16GB.

Comparison

The strongest video model under Apache-2.0, but squeezed between its own newer version and the LTX line: A14B's MoE rotates two experts, and on 16GB the peak is more alarming than nominal; what really suits 16GB is its TI2V-5B branch (720p, single model).