2025-07-28 - https://github.com/Wan-Video/Wan2.2

Wan2.2 is an advanced open-source video generative model that builds on Wan2.1 with significant improvements in architecture, training data, and compression efficiency. It introduces a Mixture-of-Experts (MoE) design that increases model capacity without added computational cost, enabling better video generation quality. Wan2.2 also incorporates cinematic-level aesthetics with detailed labels for lighting and composition and supports complex motion generation thanks to a larger dataset. The release includes three model types for text-to-video, image-to-video, and text-image-to-video generation at up to 720P and 24fps, capable of running on consumer GPUs like the RTX 4090. The 5B TI2V model uses a high-compression VAE achieving high-quality video with a compression ratio of 64 while maintaining fast generation speed. Wan2.2 outperforms state-of-the-art commercial video generation models in multiple evaluation benchmarks. The repository provides code for single and multi-GPU inference, prompt extension options, integration with Hugging Face Diffusers, and ComfyUI, making it accessible for industrial and academic use. The models are under Apache 2.0 license and the project encourages community use and citations of their 2025 arXiv publication.