Wan2.2 routes video denoising through a panel of experts

What it does Wan2.2 is a family of open-weight video foundation models from Alibaba’s Wan team. It generates video from text, images, speech, or combined prompts, shipping both a 14B-parameter MoE flagship and a 5B hybrid model that targets 720P output at 24fps. The repo provides inference code, model weights, and ready-made integrations for ComfyUI and Diffusers.

The interesting bit Most video diffusion models scale by making every layer wider; Wan2.2 instead uses a Mixture-of-Experts architecture to hand different denoising timesteps to different sub-networks. This grows total model capacity without linearly increasing per-step compute—an accounting trick that is rare in open video generation.

Key highlights

• Timestep-specialized MoE experts enlarge the model’s effective capacity while keeping the computational cost per denoising step flat.
• A 5B-parameter TI2V variant with a 16×16×4 VAE runs 720P@24fps on consumer hardware like an RTX 4090.
• Training data grew substantially over Wan2.1—65.6% more images and 83.2% more videos—with curated aesthetic labels for lighting, composition, and color tone.
• Modalities beyond vanilla text-to-video include speech-driven generation (S2V-14B) and character animation/replacement (Animate-14B).
• Broad ecosystem hooks: native ComfyUI and Diffusers support, plus community acceleration frameworks.

Caveats

• The 14B text-to-video model requires at least 80GB of VRAM for single-GPU inference, so the “consumer-grade” label only fits the smaller 5B model.
• The README asserts top performance against open and closed competitors but does not publish the comparison numbers or benchmarks.

Verdict Grab this if you want open video generation with an architectural idea beyond “stack more layers,” and you have the VRAM (or patience for offloading) to match. If you need a guaranteed drop-in replacement for proprietary APIs, the missing benchmark tables make that gamble hard to justify.