InfiniteTalk — генерация говорящего видео с синхронизацией губ

★ 6.5 · content

infinitetalk is a Claude Code skill that drives audio-synchronized talking-head video generation in both Image-to-Video (single portrait photo) and Video-to-Video (re-dubbing an existing clip) modes. It aligns lip shape, head motion, body pose, and facial expression precisely with the input audio, while streaming mode removes any hard limit on output length. A built-in TTS engine powered by Kokoro-82M lets you pass plain text instead of a pre-recorded audio file, with two-speaker dialogue supported via (s1) and (s2) markers. Inference runs through scripts/infer_infinitetalk.py with controls for resolution (480P or 720P), audio guidance scale, int8 quantization, and model offloading for GPUs below 16 GB VRAM. Model weights total roughly 30 GB; a 16 GB+ GPU and 32 GB of system RAM are recommended — making this skill the right choice for digital avatar creation, video dubbing, and virtual presenter workflows.