MiniMax-H3 Reference Adapter

The writer nodes in the pytraveler MiniMax-H3 pack take references one per socket. That's not an accident - it's what powers the drag-to-order strips and the checkbox-on-every-row rig, where each asset gets its own square, its own role and its own switch. One value per slot is the shape everything downstream is built around.

Plenty of nodes don't hand things over that way. An image batch arrives as one tensor holding many frames. A directory loader, or some other pack's output, hands you a list. A reference bundle is a whole structure in one value. None of those has a slot shape to go into, which is exactly the gap MiniMax-H3 Reference Adapter fills: references collected in together, separated back out one to a socket.

What it does

Nothing, is the honest answer - and that's the point. No model is loaded, no frame is decoded, nothing is described. Values pass through as they arrived; the node only sorts them by kind and spreads them out. A run costs about 14 milliseconds. It's a glorified routing hub, and glorified is fine because routing is precisely the fiddly bit.

Two ways in:

• items - references arriving together. The socket takes any type, because the nodes that produce collections mostly don't declare one. What each item is gets worked out from the value itself rather than from the wire, so anything that isn't an image, a clip or a sound is skipped and counted on the summary.
• bundle - a reference bundle from another pack (the node's own H3_REFS type), if you have one. Its pictures, clips and sounds are read out and placed ahead of anything on items, and the audio tracks that come with clips are treated as sounds in their own right.

Out the other side: picture_1 through picture_9, video_1 through video_3, audio_1 through audio_3, and a summary string. Nine pictures, three clips and three sounds is the room Ref2VA can hold, so that's the room there is - anything past it is reported on summary rather than silently dropped. Wire those outputs into the pack's writer and caption nodes the way you'd wire any individual asset, and the strips, numbering and checkboxes all keep working as if you'd plugged each one in by hand.

The one widget that's a real choice

split_batches decides what an image batch means, and the difference is genuine. On (the default), six frames become six references - six things the video reuses, each described and numbered separately. Off, they stay one reference made of several frames, described once - which is what a clip is. Turn it off when the batch is frames of a single shot and you don't want the writer describing the same scene six times.

Left-over outputs are harmless either way. An empty socket hands on nothing, and the writer nodes skip empties already, so plug in all nine and let the ones you didn't fill sit idle.

Why it exists as its own node

Because it can't be a socket on a writer - and this is the interesting bit. Receiving a real ComfyUI list means declaring is_input_list, and that flag isn't per-input. It rewrites the shape of every argument the node receives. Bolted onto a writer, it would change how the prompt, the duration and the options arrive. So it lives here, on a node that has nothing else to lose by it. Knowing that saves you from filing a confused feature request later: the adapter is the correct home for this job by construction.

Install

Same pack, same story as every node here: ComfyUI Manager, searching MiniMax-H3-Prompt-Rewriter-ComfyUI, or

cd ComfyUI/custom_nodes
git clone https://github.com/pytraveler/MiniMax-H3-Prompt-Rewriter-ComfyUI

then restart. This node ignores the pack's heavy requirements.txt entirely - no transformers, no bitsandbytes, no model of any kind is touched. If a reference isn't turning up where you expected, look at the line under the widget: it's the same text as the summary output, so what came in and where it went is readable off the node without wiring a preview to anything. That line will also name anything skipped or over capacity - the fix is usually one flip of split_batches.