Alibaba’s Wan team has released Wan2.2-Animate-14B, an open-weights video model that transfers a performance from a driving clip to a character image or replaces a performer in existing footage. Combining both tasks in one workflow reduces model switching and manual compositing, while ComfyUI and command-line support allow local inference.

Hugging Face reports more than 7.9 million downloads for the broader Comfy-Org package. That repository contains multiple Wan 2.2 files, and its download counter does not identify distinct users or Animate-only installations.

Two modes share the core model

Wan2.2-Animate-14B accepts a reference character image and a driving video, then runs one of two workflows:

Replacement mode includes a relighting LoRA that helps the inserted character match the source scene. The model supports 480p and 720p output, although resolution, frame count, and clip length all increase memory use.

Apache 2.0 governs the released model assets. Rights to reference art, source footage, voices, trademarks, and personal likenesses remain separate, so production systems need consent and provenance controls for their inputs and outputs.

Identity depends on preparation

Wan2.2-Animate-14B is designed to preserve the reference character’s face, body shape, clothing, and silhouette while copying the driver’s performance. Earlier animation pipelines often allowed those features to drift between frames because identity preservation competed with unconstrained video generation.