Nodes/RunningHub MiniMax H3/RunningHub MiniMax H3 Ref2VA Image Reference (Legacy)
ComfyUI Node

RunningHub MiniMax H3 Ref2VA Image Reference (Legacy)

Handing MiniMax H3 a Face (or a Product) to Keep

By RH-RunningHub·Created 29 days ago·Updated 18 days ago· 1
RunningHub MiniMax H3 Ref2VA Image Reference (Legacy)
  • image
  • references
  • references

This is the node that makes MiniMax H3's reference-based generation possible: it takes one image and hands it to the Ref2VA pipeline as a reference. In H3's omni-modal world a "reference" isn't a style hint - it's actual content the model has to honor. Give it a person and the model keeps their face; give it a product and the product stays recognizable. This node is the first brick of a reference stack.

Inputs and the chain trick

  • image - required, the IMAGE reference.
  • references - optional, a MINIMAX_H3_REFERENCES object. This is the chain trick: connect a previous reference node's output here and this node appends its image to the running list, returning a new combined references object. So you can build an ordered reference stack:
Image Reference → Audio Reference → Video Reference → Ref2VA Target → Ref2VA Encode → Sampler

Output: references - the growing MINIMAX_H3_REFERENCES list.

Why the order matters

Order is part of the contract. The Ref2VA target node and the encode node fingerprint the reference order, and H3 consumes references as an ordered multimodal context - so image-then-audio and audio-then-image are different requests, not the same thing in a different bucket. If your graph shows a mismatch error between the target and the encode node, the first thing to check is whether the reference chain order changed between runs.

The quality knob lives downstream

This node itself has no size controls - the ref_image_size option on RHMiniMaxH3Ref2VAEncode handles resolution strategy (match scales reference images to the generation canvas; max runs a separate 2048-short-edge pipeline for max fidelity at a speed cost, because every reference token participates in attention every step). For a face you really need to keep, max is worth the wait.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-MiniMax-H3.git
pip install -r ComfyUI-RH-MiniMax-H3/requirements.txt

Restart, and pull the ~95 GiB INT8 ConvRot bundle into ComfyUI/models/MiniMax-H3-INT8-CONVROT/ (hf download Gluttony10/MiniMax-H3-INT8-CONVROT --local-dir ./models/MiniMax-H3-INT8-CONVROT, ModelScope in China). Legacy node - the modern RHMiniMaxH3RefGen wraps the whole reference workflow, but if you want to understand the ordering and fingerprinting it does, this is the ground truth. License as pack-wide: H3's Community License excludes the US, EU, UK and South Korea from local-weights use.

CategoryRunningHub/MiniMax H3/ref2va

Inputs (2)

NameTypeDefaultDescription
imageIMAGE
referencesoptMINIMAX_H3_REFERENCES

Outputs (1)

NameTypeDescription
referencesMINIMAX_H3_REFERENCES