RunningHub MiniMax H3 Ref2VA Model Loader (Direct) (Legacy)

If you want to point MiniMax H3 at an image, an audio clip, or a video and have it generate around that reference - keep a character's face, match a voice, re-voice a scene - you want the Ref2VA model. This legacy loader is the Ref2VA half of H3's two DiTs, hard-wired to parse the REF2VA partition of the weights. It's the "Legacy" version, and the modern RHMiniMaxH3ModelLoader supersedes it, but if your workflow already uses it, it keeps working exactly as designed.

Ref2VA is the fancier of H3's two generation paths. FL2VA takes a frame or two; Ref2VA takes references - and because H3 is omni-modal, those references can be images, audio, or video, in any order. That's how you get the motion-transfer behavior the launch threads compared to Kling, and how you get audio that matches a given voice or room tone. This loader is just the door; the RHMiniMaxH3Ref2VAImageReference / AudioReference nodes build the reference list it feeds.

Inputs

Identical shape to its FL2VA twin, just pointed at the other partition:

• model_root - weights root, models/MiniMax-H3-INT8-CONVROT (legacy models/MiniMax-H3 still accepted). All three components must come from the same root.
• dtype - auto, bfloat16, or float16; auto/bf16 recommended.
• transformer_path - the explicit DiT file, defaulting to MiniMax-H3-Ref2VA-int8_convrot.safetensors. No silent auto-swapping.

Optional attention_backend: auto, sdpa, sage, or ck (Comfy Kitchen INT8). The sage option is a genuine win if you have SageAttention - reference tokens sit in attention for every sampling step, so attention speed matters more here than anywhere else in this pack.

Output: h3_model, ready for the Dual Sigma Sampler. Note there's no LoRA attachment on this legacy loader - that's the modern node's trick.

When to use it

Ref2VA workflows are where this pack shines, but you don't need this specific node for them anymore. If you're starting fresh, RHMiniMaxH3ModelLoader with MiniMax-H3-Ref2VA-int8_convrot.safetensors selected does the same job plus LoRA support. Keep this one around for old saved graphs - that's what legacy means here, and the pack's migration story is built around old graphs resolving cleanly.

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/. Ref2VA specifically needs system ffmpeg and ffprobe on the PATH for preparing video/audio references - the README calls that out and it's a real requirement, not decoration. And the license reminder that applies pack-wide: MiniMax H3's Community License excludes the US, EU, UK and South Korea from running the local weights.