tlemangen/freetalkdiff

FreeTalkDiff generates a talking-face video from an identity video and a motion-driving video. The inference script reads and aligns both videos, runs the generation pipeline, enhances generated faces, composites them back into the original frames, and writes the result. The pipeline combines identity and structure embeddings with inpainting diffusion, then applies temporal noise reduction. Masking and face parsing support alignment/recovery; other documented components are shown where their role is supported. Some pipeline internals are unsampled, so uncertain call relationships are omitted.