XB-BOX - 🆕 InfiniteTalk 无限接力点 (多图)

The basic InfiniteTalk relay always uses the same character - the bus's global start_image - for every segment. Fine for a monologue, useless for a story where the shot changes. XB_WanInfiniteTalk_RelayNode_MultiRef ("XB-BOX - 🆕 InfiniteTalk 无限接力点 (多图)") is the multi-image version: it lets each relay pull its own reference image straight from ComfyUI's input folder, so different segments can feature different faces, outfits or compositions while staying inside the same infinite talk pipeline.

The author's source comment calls it out as "类似 Animate 接力点" - similar to the Animate relay pattern, where a per-segment reference image is the norm. The per-segment relay recipe is unchanged (encode → sample → decode → trim → accumulate); what's new is where the character comes from.

How it works

Two new inputs do the work:

• use_local_ref_image - 继承总线全局图 (inherit the bus's global image, the default) or 独立参考图 (independent reference image). Pick independent and the relay uses your local image for this segment instead of the bus's.
• ref_image_file - a dropdown of files in ComfyUI's input directory, marked as an image-upload field. Put the reference in ComfyUI/input (or upload it through the widget) and it appears in the list. The dropdown is populated at graph-load time, so new files may need a node refresh/reload to show up.

The rest is the standard New-relay surface: positive_prompt, segment_length (81), motion_frame_count (9), relay_count (loop N segments in one node), total_frames_display, and the usual prev_video/prev_audio/audio optionals. In bus-audio mode it still respects the bus's total-frame budget and auto-stops.

Why you'd use it

Anywhere the talker changes mid-scene. A duet where each character gets their own reference and their own segment. An interview with cutaway reaction shots - character A's relay uses ref A, character B's relay uses ref B, and the bus carries the shared audio/video state between them. It's the node that turns "an endless talking head" into "an endless scene with multiple characters," without leaving the pack's bus architecture.

Install & honest 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.

The caveat is the one every multi-ref workflow shares: each new reference is a chance for identity drift. The pipeline handles continuity well, but a character who only appears from a single reference image is going to be less stable than one who's been anchored across every segment. And remember the ref_image_file list is populated from input/ at load time - drop your files there first, then refresh the node, or the dropdown will look empty.