Wan 2.2 Video ExtendExtender video

At the heart of Wan 2.2 is a sophisticated MoE transformer design. This allows the model to scale up its parameters significantly while keeping inference costs low. Different parts of the network specialize in different tasks—some for texture, some for motion, some for semantic understanding—resulting in a smarter, more capable AI.

While text-to-video is powerful, video-to-video (extension) opens new doors. Users can provide a starting video, and the model acts as a virtual cinematographer, continuing the shot. This is crucial for fixing clips that ended too soon or for stitching together a cohesive story from short segments.

You aren't limited to just 'more of the same'. You can use text prompts to steer the extension. For example, if your video shows a car driving, you can prompt 'the car turns onto a dirt road', and the extension will attempt to follow that narrative thread while maintaining visual consistency.

Wan 2.2 supports mainstream resolutions like 720p and 1080p. It uses a robust 3D VAE for video encoding and decoding, ensuring that the generated frames are crisp, artifact-free, and temporally stable. Say goodbye to flickering backgrounds and morphing objects.

From fast-paced action sequences to slow, emotional close-ups, Wan 2.2 handles temporal dynamics with grace. It understands the physics of motion, ensuring that gravity, momentum, and biological movements look natural and convincing.

Trained on a massive dataset of high-quality videos and images, the model has a vast internal library of visual concepts. This allows it to generate realistic extensions for a wide variety of subjects, from photorealistic humans to stylized anime and abstract art.