What models and nodes are essential for this to work?

At minimum: UNETLoader with Wan2.1 I2V 14B 480p weights, VAELoader, CLIPLoader + CLIPTextEncode (for camera/scene prompts), AudioEncoderLoader + AudioEncoderEncode (for voice features), LoadImage for the start frame, WanInfiniteTalkToVideo to drive motion, VAEDecode to get frames, and CreateVideo/SaveVideo to export. LoraLoaderModelOnly and ModelPatchLoader are optional for style or InfiniteTalk patches.

How do I match the video length to my audio?

Use the Video Extend group. Each copy adds 81 frames at 25 fps (~3.24 s). Compute ceil(audio_seconds / 3.24) and duplicate the group that many times (box‑select, Ctrl‑C, then Ctrl‑Shift‑V). If you have multiple audio clips, assemble them with AudioConcat first so the total duration is clear.

Why does the background or identity drift?

Over‑strong text or style conditioning can push the look away from the source. Reduce CLIP influence using ConditioningZeroOut, avoid heavy style LoRAs (or lower their weight in LoraLoaderModelOnly), and keep prompts descriptive (camera, staging) rather than stylistic. Also ensure masks are tight around speaking characters so motion doesn’t bleed into the background.

What audio format works best and how do I improve lip sync?

Use clean, single‑speaker audio per track where possible. Common WAV/MP3 files load via LoadAudio; if a file fails, convert to a standard WAV (e.g., 16–48 kHz, mono). Better lip sync comes from clear diction and minimal background noise. For multi‑speaker scenes, keep separate audio files and masks per character and arrange them with AudioConcat.

InfiniteTalk: Audio-Driven Full-Body Video Dubbing

InfiniteTalk: Audio-Driven Full-Body Video Dubbing turns a single image and one or more audio tracks into a lip‑synced, full‑body video while preserving the original identity, background, and camera motion. The core of the pipeline is WanInfiniteTalkToVideo, which fuses features from the loaded Wan2.1 I2V model with an audio embedding to drive body and facial motion. Step1 - Load Models prepares the UNet and VAE (UNETLoader, VAELoader) for image‑to‑video synthesis, and loads text and audio backbones (CLIPLoader with CLIPTextEncode for camera/stage prompts, and AudioEncoderLoader with AudioEncoderEncode for voice features). You import the starting frame via LoadImage, then draw masks per character in the MaskEditor to localize motion to each speaker.

FAQ

Sık Sorulan Sorular