AI Character Replacement Workflows

The real worry with any ai character replacement job is that the new character ends up pasted on top of the old one: the face changes but the body drifts, the motion goes rubbery, or the lighting no longer matches the scene. Most consumer tools that promise ai character replacement only retouch the head and leave the rest of the shot fighting itself. The Wan 2.2 Animate workflows on this page work the other way around. They read the full performance out of a driving video, body pose and facial expression together, and rebuild a new character performing that exact motion, so the whole figure moves as one piece rather than a swapped face on borrowed shoulders.

Wan 2.2 Animate is a 14B open-weights model from Alibaba's Tongyi Lab, released under Apache 2.0, and it runs two jobs from the same node graph. In replacement, or Mix, mode it substitutes the character in your existing footage while keeping the original background, so a person, an anime figure, or a 3D render steps into the scene and the room stays put. In animation, or Move, mode you give it one clean reference image and a driving clip, and it animates that still character with the performer's motion on the reference image's own backdrop.

What makes the output hold together is how the model reads motion. Wan Animate uses spatially aligned skeleton signals to copy body movement and implicit facial features to drive expression, so gestures, head turns, and lip articulation land on the new character in the right frames. A Relighting LoRA matches the lighting and color of the target scene to the character it inserts, the step that stops a swap from looking cut out. Because the pipeline tracks pose and face directly, you animate a character from a reference video without the rotoscoping, masking, or hand-painted keyframes that traditional character animation and replacement used to demand.

Everything here is an open, editable ComfyUI graph, so the controls are yours. Run it locally on your own GPU or on Comfy Cloud with no install, pick the FP8 or BF16 build to fit your VRAM, and add the 4-step acceleration LoRA for a faster preview. Clips extend in segments of roughly 77 frames, near five seconds each, and chain for longer takes. Treat the stack as ai motion transfer you fully own: change the reference character, retime the driving clip, adjust the relight, and re-run any node until the character swap in video reads as a real performance instead of an edit.