MiniMax H3 CLIP Name

MiniMaxH3CLIPName is the quietest node in this pack, and you might not need it at all - until your workflow has more than one cached H3 node and you've picked the encoder checkpoint twice, set them slightly differently, and spent an hour wondering why two branches of the graph don't match. It exists to answer one failure mode: repetition drift.

The cached FL2VA and Ref2VA nodes in ComfyUI-MiniMaxH3-CLIPCached take their encoder as a clip_name dropdown - the Qwen3-VL checkpoint filename from models/text_encoders - rather than a loaded CLIP object. If a workflow uses two of those nodes, you're selecting the same checkpoint in two places. That's fine until it isn't: change the encoder once and miss the other dropdown, and your cache fingerprints diverge silently because the checkpoint identity is part of the cache key. This node is the standard ComfyUI answer to "one value, many consumers": a single source you fan out from.

What it actually does

Nothing fancy. It's a one-widget value node carrying the exact same encoder dropdown as the cached FL2VA/Ref2VA nodes, and it re-emits the selected filename as a clip_name output (typed as a COMBO so it wires into the nodes' dropdown inputs cleanly). Crucially, it does not load Qwen3-VL, does not generate conditioning, and does not create a cache entry. It's a string pass-through. There's no model-loading trap lurking here - that's the whole point.

To use it: add the node, pick the checkpoint once, then on each cached FL2VA/Ref2VA node right-click the clip_name widget and choose Convert widget to input, and wire this node's clip_name output into the new socket. Now the encoder is chosen in exactly one place, and swapping it updates every downstream node at once. It shares the four cached nodes' "one encoder, four conditioning nodes" house style, so treat it as the shared control for FL2VA, Ref2VA, and both Dual Resolution variants.

Installing

Same pack as the rest - ComfyUI Manager (search "MiniMax H3 CLIP-Cached") or a clone:

cd ComfyUI/custom_nodes
git clone https://github.com/Mu5hr00moO/ComfyUI-MiniMaxH3-CLIPCached

Restart ComfyUI. The node shows up under model/conditioning/minimax/cached alongside the others. Like the rest of the pack it assumes ComfyUI v0.30.0+ with native MiniMax H3 support, and it's only useful if you've already got the H3 encoder checkpoint in models/text_encoders.

The one trap

Don't mistake the output for a loaded CLIP. Somebody wiring this into a normal CLIP Text Encode node would be chasing the wrong tool entirely - this emits a checkpoint name, not a model, and it only feeds the clip_name inputs of this pack's nodes. Beyond that there's genuinely little to get wrong: if the dropdown looks empty, you're missing the encoder file in models/text_encoders, and no helper node will fix that. When you only run a single cached H3 node, skip this and set clip_name on the node itself; the helper earns its keep the moment the second encoder dropdown appears.