H3 Text Multiline (QQ)

MiniMax H3 is the 33B omni-modal video model that opened its weights in August 2026 - text, image, video and audio all live in one prompt context, and you point the model at specific media by writing tags like <Picture 1> into the text. H3TextMultiline is a ComfyUI port of the "H3 Easy" frontend's way of handling that: you type @Picture 1 in a normal box, the tags get resolved to the real H3 format for you, and reordering your media never breaks the prompt.

That last part is the whole point. Hand-writing <Picture 1> means keeping the numbering in sync with the order you wired media in, and the first time you drag a wire to a different socket every tag downstream is wrong. This node counts your media in connection order and does the renumbering at execution time.

How it works

You type into a multiline text box and reference media with @ mentions: @Picture 1, @Video 2, @Audio 1 - @Image is accepted as a synonym for Picture. On the canvas, the node's frontend treats anything connected to it as virtual media wires: images, video, and audio you hook up (from a loader, from a VACE-style reference, wherever). The JS collects them in order into a hidden media_order widget - something like "image,video,image" - and at execution the backend walks your text, matches each mention against that order, and swaps in the official tag: @Picture 2 becomes <Picture 2> only if the second connected image is really second. No match for a mention, and the @ text just stays visible instead of silently producing a wrong tag.

There's also an @ mention popup over the text widget so you can pick from connected media instead of typing the ordinal from memory. The media input on the node is a wildcard anchor - the real wiring lives in the node's saved properties, so it round-trips with the workflow file.

The inputs that matter:

• text (multiline STRING) - your prompt with @ mentions. This is the one you live in.
• mode (reference default, or image) - reference mode allows up to 9 images, 3 videos and 3 audio (15 media total); image mode caps at 2 images and drops anything else. That's the H3 Easy limit scheme, ported verbatim.
• strip_newlines (BOOLEAN) - collapses line breaks, handy when the model wants a single-line prompt.

Output is one text STRING, which feeds whatever consumes your H3 prompt.

Ground rules and gotchas

First, the model itself: H3's community license excludes the US, EU, UK and South Korea from running the local weights. This node can't fix that - if you're in a restricted region, only the hosted API is legitimately available. Second, it's a prompt-builder, not a runner: this pack is Wan-focused, and H3TextMultiline is the odd node out, built to feed an H3 Easy pipeline rather than run H3 itself.

Install

Get it via ComfyUI Manager (search "WanVideoWrapper QQ", now renamed SA-Nodes-QQ) or:

cd ComfyUI/custom_nodes/
git clone https://github.com/siraxe/ComfyUI-WanVideoWrapper_QQ.git

Restart ComfyUI. No extra dependencies beyond ComfyUI - no requirements.txt, no model downloads for this node. Watch out for the repo rename: v1.3.4 moved it to ComfyUI-SA-Nodes-QQ, so if older workflows trigger a spurious install, delete the stale wanwrapper_qq folder from custom_nodes.

Brand-new and experimental, so expect rough edges. But if you're building H3 prompts with multiple references, it's the first node that makes the tag bookkeeping someone else's problem.