XB-BOX - 🆕 InfiniteTalk 无限接力点 (全能)

If the pack's InfiniteTalk relays are a family, this is the one with every option bolted on. XB_WanInfiniteTalk_RelayNode_AllInOne ("XB-BOX - 🆕 InfiniteTalk 无限接力点 (全能)") is the "omnipotent" variant: it can start a segment from the bus's global image, from an independent reference file, or from the previous segment's last frame - whichever you need, selected per node from a single start_image_mode dropdown.

The author's comment in the source spells out the design: "首帧来源三选一" - the first-frame source is three choices - plus automatic alignment of audio cropping and overlap frames so the trim and length stay consistent. Everything else is the full New-relay feature set: relay_count internal looping, motion_frame_count transition overlap, total_frames_display, bus-audio support, and audio/video accumulation.

The three start modes

• 继承总线全局图 (inherit bus global image) - the default; use the character everyone shares. The standard talking-head mode.
• 独立参考图 (independent reference image) - pick a local file via ref_image_file (dropdown populated from ComfyUI's input folder, with upload support). For a segment that features a different subject.
• 继承前段尾帧 (inherit previous segment's last frame) - the true "continuous take" mode: each segment starts exactly where the last one ended, which is the strongest continuity you can get short of generating the whole thing in one pass. This is the mode for "one long uninterrupted conversation" where you don't want the character re-anchored to a still.

Because the tooltip on motion_frame_count notes audio and video cropping auto-align, mode 3 works with audio sync preserved - the overlap frames you trim off the head are matched on both tracks.

Inputs

wan_infinitetalk_bus, positive_prompt, segment_length (81), start_image_mode, ref_image_file, motion_frame_count (9), relay_count (1), total_frames_display, plus optional prev_video, prev_audio, audio. Outputs: the bus, 🎞️ 累加视频流, 🔊 累加音频流.

When to reach for it

This is the node to start with if you're not sure which flavor you need - it's strictly more capable than the other relay variants, and its defaults (global image, 9-frame overlap) match the boring single-talker case. The 继承前段尾帧 mode is genuinely the smoothest option for seamless long talking takes, since it never snaps the character back to a still.

Install & caveats

Pack standard: ComfyUI Manager → XB_ToolBox, or git clone https://github.com/wjluoxiao/XB_ToolBox.git into custom_nodes, restart. Deps opencv-python + easyocr, no weights bundled - InfiniteTalk weights, audio encoder, VAE/CLIP/T5 are yours.

Honest read: "all-in-one" also means "more state to get wrong." If a segment isn't using the image you expected, check start_image_mode before anything else - it's the one setting that silently changes the whole character source. And as always with InfiniteTalk, each relay_count step is a full 14B-class generation, so mode 3 smoothness comes at the same clock cost as everything else here.