wan-video/wan2.2 is a repository containing the official inference code and weights for the Wan2.2 model, which generates videos from text, image, and audio conditions. The README states that it provides runnable code and weights for five models: Wan2.2-T2V-A14B, Wan2.2-I2V-A14B, Wan2.2-TI2V-5B, Wan2.2-S2V-14B, and Wan2.2-Animate-14B.
Wan2.2 introduces a Mixture-of-Experts (MoE) architecture into a video diffusion model, and is a video generation model that finely controls aesthetic elements such as lighting, composition, and color tone. MoE is a structure that computes only the needed specialized modules among several. This repository was created on July 28, 2025, is written in Python, and follows the Apache-2.0 license. Its GitHub category is image/video models, and it has 18,059 stars. Public video generation models such as T2V, I2V, and TI2V are included together with their runnable code.
However, this repository has no public releases. Whether the aesthetic control described in the README translates into actual video quality cannot be known without looking at the generated results. There are no published benchmark figures, and among the score items only Momentum is filled in at 5.9, while Freshness, Technology, and Quality are empty. Judging which of the five models is better is premature based on this material alone. The fact that the official code and weights have been released publicly and the level its output reaches are separate issues.