H3 Reference Inputs

Ref2VA prompts for MiniMax H3 are written in reference tokens - <Picture 1>, <Video 1>, <Audio 1> - and the whole game is keeping those tokens straight with the media you actually loaded. H3 Reference Inputs is the companion node that makes that painless: instead of hand-typing token names and hoping they match what's on your graph, you plug your media into labeled sockets that grow themselves, and the editor sees the names.

How it works. This is an authoring node, not a generator. It takes up to nine reference pictures, three video frame batches, and six audio connections, and packs them into a single references bundle (a custom H3_PROMPT_REFERENCES type) that you feed into H3 Prompt IDE. The sockets are "Autogrow" style under ComfyUI's V3 node API - plug in <Picture 1> and <Picture 2> appears, and so on up to the H3 limits. The backend preserves H3's presentation order even if your workflow serializer shuffles the insertion order, which sounds pedantic until you've had a prompt silently come out with the wrong picture mapping.

The three inputs, plainly:

• pictures - up to 9 images, numbered <Picture 1>–<Picture 9> in H3 presentation order.
• videos - up to 3 IMAGE frame batches, <Video 1>–<Video 3>. Use the exact same frame batches you'd feed the native H3 ref_videos input.
• audios - up to 6 labels, <Audio 1>–<Audio 6>, but the order matters: connected video soundtracks first, then standalone audio. That's H3's native presentation order; get it backwards and your audio tokens won't line up.

The single output, references, goes into the editor's references input. There it supplies the token palette and miniatures so you can see which thumbnail is <Picture 3> while you're writing - genuinely useful when a shot list has six pictures and you've lost track of which is which. When you queue text only, the media isn't evaluated; it's an authoring aid until the native H3 conditioner does the real work. For actual Ref2VA generation, you still connect the same loaders to the matching inputs on the native H3 conditioner - this node never replaces that, it just keeps the editor honest about names.

Install and gotchas. Same pack, same steps as the editor - ComfyUI Manager (search "H3 Prompt IDE") or git clone https://github.com/ethanfel/ComfyUI-H3-Prompt-IDE.git into custom_nodes, restart, find both nodes under text → H3 Prompt IDE. No extra dependencies, no model files. Requires a current ComfyUI with the V3 API and Autogrow.

One deliberate omission worth knowing: this pack intentionally does not include the Motion Context @tag / @@@@tag compilation from the parent Contex-Loop pack. If your workflow depends on those tags, this standalone editor isn't the one - it's built for a normal text-output pipeline. And don't expect it to condition the model by itself; it's half of a two-node pair, and the other half is the editor it feeds.