Viggle-Animate: Replace Video Characters in 26 Seconds Without Pose Extraction

Replace the character in a single frame using an image editor, and the system naturally composites that character across all remaining frames. By bypassing intermediate steps such as pose estimators or segmentation masks used in existing methods, the original motion and camera work are preserved exactly.

Rendering 124 frames takes 26 seconds on a B200 GPU, which is 6.1 times faster than Wan2.2-Animate-14B. Inference completes with only 3 forward passes thanks to DMD distillation, and no text encoder or separate model loading is required, eliminating orchestration overhead.

Non-human animals, robots, and stylized characters can also be animated as long as a painted reference image is provided. However, there are still limitations with complex multi-character scenes or precise lip-syncing; providing an audio track alongside the input improves lip-sync accuracy.