XB-BOX - 🎵 InfiniteTalk 无限对口型总线

InfiniteTalk is the Wan-ecosystem trick for talking-head video that doesn't loop or freeze: an audio-conditioned model that drives a character's lips and motion from a voice track, segment by segment, so the clip can run as long as the audio does. XB_WanInfiniteTalk_ParamBus ("XB-BOX - 🎵 InfiniteTalk 无限对口型总线") is the pack's command center for that pipeline - the one place you set every shared parameter, load the models, and hand a single WAN_INFINITETALK_BUS to the relay nodes that actually generate each segment.

If you've used the pack's plain XB_Wan_ParamBus, this is its audio-aware sibling. Everything the Wan bus does for I2V relays, this does for InfiniteTalk relays - plus the model plumbing and the audio handling those relays need, which is why it takes more inputs.

How it works

On the model side it accepts the full stack: model and model_patch (the audio-conditioned checkpoint plus its patch - the source clones and patches a copy so an interrupted run can't poison the base model), clip, vae, audio_encoder, and clip_vision. Then the usual generation parameters: width/height (832×480 default), fps, negative_prompt, steps (default 4 - the distilled regime), cfg (1.0), sampler/scheduler/seed, plus the VAE chunking knobs (vae_encode_tile_size, vae_decode_tile_size, spatial_overlap, temporal_chunk_size, temporal_overlap) and a cleanup dropdown for post-run VRAM nuking.

Then the InfiniteTalk-specific pieces:

• total_frames - per its tooltip, this is "bus audio mode": the total frame count from your audio-slicing node, which the relays consume until it's exhausted and then auto-stop. 0 means "independent audio mode," where each relay feeds its own audio.
• audio (optional) - in bus audio mode you drop your long audio here and the bus encodes it once; relays then slice it automatically by frame offset. One encode for the whole track instead of re-encoding per segment.
• motion_frame_count (default 9) and audio_scale (default 1.0) - the overlap for seamless transitions between segments, and how strongly the audio drives the motion.
• start_image (optional) - the character's face; if you also wire clip_vision, the bus encodes it into clip-vision conditioning for identity retention.

The single output is the WAN_INFINITETALK_BUS, which you pass to any of the pack's InfiniteTalk relay nodes.

What to set

For a first run: models + clip_vision + start_image + negative_prompt, and pick your audio mode. Bus mode (long audio + total_frames from the slicer) is the one that makes "one video as long as the podcast" a real workflow. Independent mode is for per-segment audio, which is easier to iterate on. motion_frame_count of 9 is a sane overlap - lower it if transitions feel heavy, raise it if they feel jumpy.

Install

ComfyUI Manager → XB_ToolBox, or git clone https://github.com/wjluoxiao/XB_ToolBox.git into custom_nodes, restart. Declared deps: opencv-python, easyocr; no weights ship. You need the InfiniteTalk weights, the Wan audio encoder (the AUDIO_ENCODER this node takes is loaded by ComfyUI's audio-encoder node), a Wan VAE/CLIP/T5, and a CLIP Vision for identity. InfiniteTalk is a real, working open model - community threads report seamless 20-second takes with no visible transitions - but it's a niche corner of the ecosystem, so expect to debug the plumbing yourself.