MiniMax H3 VLM System Instruction Presets

MiniMax H3 is the 33B omni-modal video model whose text conditioning runs through a Qwen3-VL 32B encoder - and like most modern video VLMs, it's picky about its system instructions. UC_MiniMaxH3VLMSysInstrPresets is a dropdown of curated system-instruction strings built for H3, so you don't have to hand-write the boilerplate that tells the encoder how to behave. Pick a preset, get a STRING, feed it where your H3 conditioning path expects a system instruction.

It's part of the MiniMax H3 tooling in silveroxides/ComfyUI-UtilsCollection, which is one of the few packs that shipped day-zero-style support for H3's ComfyUI integration. The presets cover the situations people actually hit: a base minimax_h3_base, frame-anchor variants like minimax_h3_first_last_frame and minimax_h3_last_frame, reference-image variants (minimax_h3_reference, minimax_h3_full_reference), and a batch of style-specific reference presets (paper collage, stop-motion, product ad, and so on). There are also timeline-oriented presets like minimax_h3_timeline_fl2va and minimax_h3_timeline_ref2va for prompt-driven video structure.

What's on the node

• preset - a single dropdown. The default is minimax_h3_base, which is the right starting point for ordinary text-to-video. If you're building a first/last-frame or reference workflow, switch to the matching preset rather than improvising.
• system_instruction - the string output. Wire it into wherever your H3 encoder node takes its system instruction.

That's the entire node: one dropdown, one string out. The value is in the strings themselves - they encode the phrasing H3's encoder responds well to, which you'd otherwise be debugging through trial and error.

Installing it

Via the pack:

• ComfyUI Manager: search "ComfyUI-UtilsCollection".
• Or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart ComfyUI. No models to download for this node - the H3 models themselves are a separate download, and remember H3's community license excludes the US, EU, UK and Korea, so check that before you plan a workflow around it.

Gotchas

The preset is a starting text, not magic. It pairs with your own prompt, and H3's results still depend on how you phrase the actual video prompt - the system instruction just sets the encoder's frame of reference. Also, the experimental timeline presets live in the Experimental variants of this node (...VLMSysInstrPresetsExperimental), so if you can't find the timeline options in this node, you're looking at the wrong one. And if you need to inject your own overrides on top of a preset, that's what the Advanced variant is for - this plain version is deliberately one-dropdown simple.