Wan-Animate: Unified Character Animation and Replacement (humanaigc.github.io)

Wan-Animate is a unified framework for both character animation and character replacement built on the Wan (Wan-I2V) model. Given a single character image and a reference video, it can either animate the character to precisely reproduce the reference’s expressions and movements or insert the animated character back into the scene as a seamless replacement. The authors modify the model’s input paradigm to explicitly separate reference conditions from generation regions and to encode temporal frame guidance and environmental cues into a common symbolic representation, enabling dual-mode operation from one architecture. Technically, Wan-Animate uses spatially-aligned skeleton signals to drive body motion and implicit facial features extracted from the source image to reenact expressions, giving high controllability and expressiveness. For replacement use-cases it adds a Relighting LoRA module that adapts the character’s appearance to match scene lighting and color tone while preserving identity-consistent appearance. Experiments reportedly show state-of-the-art results, and the team will open-source model weights and code. For the AI/ML community this unification simplifies pipelines for animation/replacement, provides modular control signals (skeleton + implicit face embeddings), and introduces a light-weight relighting fine-tuning approach that could be reused in other generative-video tasks.