Alibaba Unveils Wan2.2 AI Video Generation Model

Alibaba Group has announced the release of Wan2.2, an open-source suite of AI video generation models. This new series introduces a Mixture-of-Experts (MoE) architecture, enhancing video quality and efficiency. The Wan2.2 models include text-to-video and image-to-video capabilities, allowing users to create cinematic-style videos with customizable aesthetic preferences.

The Wan2.2 series features a hybrid model, Wan2.2-TI2V-5B, which supports both text-to-video and image-to-video tasks. This model is designed for efficient deployment on consumer-grade GPUs, such as the NVIDIA RTX 4090, making high-definition video generation accessible to a broader audience. The models operate with a total of 27 billion parameters, but only 14 billion are active at any given step, maintaining computational efficiency.

Wan AI | Wan 2.2: Leading AI Video Generation Model is available under an Apache 2.0 license, reflecting Alibaba's strategy to challenge closed-source competitors by offering powerful, free tools to developers and researchers. The models are accessible on platforms like Hugging Face and GitHub, promoting community engagement and innovation in the open-source ecosystem.