RunningHub MiniMax H3 Ref2VA Encode (Legacy)

This is where MiniMax H3's reference generation actually comes together. The reference nodes just build a list of images, audio and video; this node takes that ordered list plus your prompt and runs it through H3's Qwen3-VL encoder and VAEs to produce the MINIMAX_H3_CONDITIONING the sampler denoises against. If the reference nodes are the ingredients, this is the oven.

Inputs

• h3_text_encoder - the Qwen3-VL loader handle. It's doing the heavy lifting here, fusing text and multimodal references into one conditioning context - which is exactly the "omni-modal" design H3 was built around.
• h3_vae_bundle - the dual VAE bundle, needed to encode image and video references into H3's latent space.
• target - the MINIMAX_H3_TARGET from the Ref2VA Target node. Resolution, duration, and the agreed contract for the whole graph.
• references - the MINIMAX_H3_REFERENCES list from the chained reference nodes. Order matters; this is where it gets consumed.
• prompt - your text direction, multiline. With references present, the prompt is where you describe what should happen - the action, motion, and any audio direction the references don't already dictate.

The one optional input worth knowing:

• ref_image_size - match (default) scales reference images to the generation canvas by area (downscale only); max runs the reference pipeline at an independent 2048 short edge for maximum fidelity. The tooltip is blunt about the cost: reference tokens sit in attention for every sampling step, so max can be several times slower. Use it when identity really matters - a face you need to keep - and use match otherwise.

Output: conditioning - the MINIMAX_H3_CONDITIONING feeding the Dual Sigma Sampler.

Gotchas

The encode node is where fingerprint mismatches surface. It checks that the target's reference order matches what you're encoding - change the chain order between building the target and running the encode and you'll get a hard error rather than silently wrong output. That's the pack being helpful; chase it back to the reference chain, don't fight it.

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, pull the ~95 GiB INT8 ConvRot bundle into ComfyUI/models/MiniMax-H3-INT8-CONVROT/, and have system ffmpeg/ffprobe for video/audio references. Legacy node - RHMiniMaxH3RefGen wraps this flow today - but the encode step and its ref_image_size tradeoff are the real mechanics either way. License as pack-wide: H3's Community License excludes the US, EU, UK and South Korea from local-weights use.