⭐ Star Minimax Multiref Option

The setup

MiniMax H3 is the 33B omni-modal video model that made the biggest splash of the 2026 video season: 4–15 second clips up to 2K at 24 fps, with audio generated in the same pass rather than bolted on afterwards. Before you download 33B of weights, know that the community licence excludes the US, EU, UK and South Korea from the territory where you're licensed to run them. That's about the model, not this node, but it's better learned now than after the download.

What people actually use H3 for is reference-to-video: you hand it images and clips and it maintains identity across the shot. And the interesting half of that - anchoring a specific frame at a specific moment - is what this node is about.

What it replaces

Stock ComfyUI does keyframe anchoring with a chain of core MiniMaxH3AddGuide nodes. One node per guide, and each one wants a frame index, so each one also needs a Math Expression node converting your seconds into frames. Five anchors, ten nodes, all named almost identically. It's the same story as the LTX branch of this idea, where the ability to inject frames mid-generation existed upstream from day one but the official workflows never demoed it, so the community built wrappers around a feature it had to discover.

⭐ Star Minimax Multiref Option is that chain collapsed into one node. Up to nine reference images/clips on the left, nine matching start @ guide_image_N (s) widgets on the right, one output. Same math, same code path, just not ten nodes of it.

How it works

The node is deliberately dumb: it collects each connected image slot with its start time into an {image, start_seconds} pair, logs how many guides it found, and hands the list out as a MULTIREF_SETTINGS bundle. All the actual work happens inside ⭐ Star Minimax All In One, which calls the core MiniMaxH3AddGuide.execute in-process - the same function the stock node runs - with frame_idx = round(seconds * 24), exactly what the template's Math Expression nodes compute.

Two behaviours fall out of that code path and they're the reason this node is worth understanding rather than just wiring up:

• A single image becomes a still anchor. A batch of 5+ frames becomes a clip anchor, cropped to the valid 17k+5 clip lengths.
• Guides act on their own timeline position and nothing else. They are not part of the <Picture i> reference ordering - that's the ref_images / ref_videos inputs on the all-in-one. Guides say when; references say who.

The guides are also carried into the optional latent-upscale refine pass, resolution-matched, so your anchors survive the second pass without you re-wiring anything.

Wiring it up

multiref_settings → the multiref_settings input on ⭐ Star Minimax All In One. That's the entire connection. It works in video and image mode, and the node's first image slot is the only required input - slots 1 through 8 are optional and unconnected slots simply don't produce a guide.

Timing: it's 24 fps, hard-coded, because that's H3's native rate. One second in the widget equals frame 24. Negative start times count back from the end of the video, which is the easy way to say "this frame is the last thing you see" without knowing the clip length in advance.

Installing

Same pack, same two routes. Manager: search Starnodes → Install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

This node has no Python dependency of its own and does not need the separate ComfyUI-MiniMaxH3Mod pack - that one is required by its sibling, ⭐ Star Ref Mod Option. It does need a ComfyUI new enough to contain the core MiniMaxH3AddGuide node from the MiniMax H3 multiframe template; if yours doesn't, you get an error naming that missing node rather than a silent failure. The H3 models for the all-in-one (text encoder, both VAEs) load internally from files you supply.

Where this goes wrong

Piling guides up. Every guide-anchored video workflow in the ecosystem hits the same wall: put anchors too close together and the model has to reconcile competing references, and you get flicker and morphing between them. People using the LTX equivalent report mismatch and flicker in the frames right next to reference frames. Start with a first and last anchor, add a middle one only where the shot needs it, and space them out - a handful of anchors is the normal dose for a five-second clip, not nine.

Confusing guides with references. Anchoring a face at 2.0 s doesn't make the model keep that face throughout - if your subject drifts, that's a ref_images problem. Feed fewer, cleaner reference images at high retention rather than adding timeline anchors.

Assuming zero means "the beginning of the clip". It does - and several guides set to 0.0 stack on frame 0 and fight each other. This node's console line ([Star Multiref Option] 3 guide(s) @ 0.00s, 0.00s, 2.50s) is the fastest way to catch that.