MiniMax-H3 Universal Writer

This is the everything node of the pack: it captions your references, decides their order, and writes the finished MiniMax-H3 prompt from the guide, covering all five tasks - T2VA through Ref2VA. It's what the Multi Reference Caption plus both writer nodes do, folded into a single box. Those stay exactly as they are; nothing you've built stops working.

The reason to fold them together is order, and order is genuinely load-bearing. Picture 1 and Picture 2 are not interchangeable in FL2VA: one opens the video and the other closes it. Until this node, the only thing deciding which was which was the order the caption node happened to write them in - which came in turn from which slot each was plugged into. Real, and nowhere on screen. Here, it's the whole point.

One socket, and a strip that decides everything

references is a single growing socket that takes an image, a clip or a sound - more slots appear as you fill them. There's no wrong socket to plug into, because which socket you used no longer decides anything.

Below the inputs, the strip shows every connected reference as a colored square, numbered exactly as the block will number it:

| Square | Label | What it means | |---|---|---| | pic | <Picture N> | an image as an actual frame - first, last, key, composition anchor | | subj | <Subject N> | reusable visible content - a person, an animal, a place, a style | | vid | <Video N> | a clip, or a batch of frames read as one | | aud | <Audio N> | a voice timbre, music, ambience, an effect |

• Drag a square to move it; numbering follows immediately. The number is a position, the slot name under it stays with the square.
• Click its label to change what an image is for - pic → subj → vid and round again. Clips and sounds are what they are.
• Click its number (or the checkbox on the slot's row) to switch a reference off without unplugging it.

The task switch reads the strip: a task greys out while the strip can't supply what it's written from. Nothing connected → only T2VA lit. One picture → I2VA and L2VA. Two → FL2VA. A task you picked that becomes impossible turns red rather than quietly failing later. A mismatch is refused before anything is downloaded, and the message names the strip, because that's where the fix is.

Inputs to know

• task - all five. Ref2VA uses the six-section full-reference guide and accepts any number of references (at least one).
• caption_model / writer_model - two models, two jobs. The captioner reads the references; the writer produces the prompt from the guide. Both ordinary GGUFs, no LoRA anywhere.
• duration - a slider in tenths of a second, and its range is the node's own max_duration property (right-click → Properties Panel), 30 seconds until you change it. The server accepts up to ten minutes.
• clip - connect a multimodal encoder from CLIPLoader and it replaces caption_model for reading references, staying loaded between them. On T2VA no captioner is touched at all.
• previous - a reference block from an earlier caption node; its labels are counted, so this node's references continue from where it stopped.
• max_frames / context_size - the captioner's knobs; context_size 0 sizes itself from the references and your card.

Outputs are the union of both writers' - the three T2VA fields, the six Ref2VA fields, plus reference_assets and captions. A task that doesn't write a field leaves it empty, because a node's outputs can't change with a widget value.

Install and gotchas

Same pack, same install (ComfyUI Manager → search "MiniMax-H3", or git clone into custom_nodes). Weights fetch on first use.

• Needs a recent ComfyUI - the growing references input is io.Autogrow from the v3 node API. On an older install this node goes missing while the rest of the pack registers.
• Block order is strip order, not guide order. Multi Reference Caption writes subjects-pictures-videos-audio whatever the wiring says; this one writes in strip order, because the strip is the point. An untouched strip is slot order.
• No UI fallback is a loss. If the interface script doesn't load, the strip, task switch and ratio picker fall back to plain widgets and the node still runs - but the checkboxes on the input rows are canvas-drawn, so with Modern Node Design (Nodes 2.0) you switch references off via the square's number instead.