RunningHub MiniMax H3 Ref2VA Target (Legacy)

The Ref2VA (reference-to-video-and-audio) path in MiniMax H3 is special because it can be self-describing: your references can imply the output shape and length instead of you spelling them out. RHMiniMaxH3Ref2VATarget is the legacy node that did that resolution step, and while it's deprecated now, it's the cleanest illustration of how H3's reference mode thinks.

How it works

Where the T2VA target makes you name a duration, the Ref2VA target can take the length from a reference. Set duration_seconds to 0 and the node derives the duration from a single actual audio reference - the "only one audio ref, so that's the running time" rule. The aspect ratio defaults to auto precisely because images and videos give the model something to match; that's why auto is legal here and illegal on the text path.

The other inputs are the same shape as the T2VA target: aspect_ratio (auto, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16), duration_seconds (0–15, 0 = derive), and optional width / height (0 = by aspect ratio, both-or-neither, 32-aligned). But there's a difference in what it consumes: this node takes a references input (type MINIMAX_H3_REFERENCES) - the ordered list built by the legacy Ref2VA image/audio/video reference nodes - and it fingerprints the reference order into the target so the downstream encode node can sanity-check that the media sequence still matches.

Outputs

target (MINIMAX_H3_TARGET) and shape_info (STRING) - the resolved width/height/frame count and the separate video- and audio-latent dimensions, as JSON, for debugging.

The honest recommendation

Skip it. It's deprecated, hidden from search, and kept around so old workflows don't break. RHMiniMaxH3RefGen does all of this internally: you feed it reference images, videos, and audio through autogrow ports, and it builds the sequence, resolves the target (including the duration=0-from-audio trick), and encodes - in one node. Reach for the legacy target only if you're maintaining a pre-refactor workflow. The pack's migrator can modernize it for you:

python3 tools/migrate_workflow.py old_workflow.json --in-place

Install

The standard pack routine - Manager (search "ComfyUI-RH-MiniMax-H3") or:

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, then the ~95 GiB INT8 ConvRot model bundle into models/MiniMax-H3-INT8-CONVROT/. One thing that carries over from the legacy chain and still bites people on the modern node: if duration=0 gives you an unexpected length, it's because there was more than one audio reference and the derivation rule didn't apply - in that case you must set the duration explicitly.