RunningHub MiniMax H3 FL2VA Model Loader (Direct) (Legacy)

MiniMax H3 comes in two DiT flavors: FL2VA, the first/last-frame keyframe model, and Ref2VA, the multimodal-reference model. This legacy loader is the FL2VA half, hard-wired so it always parses the FL2VA partition of the weights. It's a legacy node - the pack now steers you to the single RHMiniMaxH3ModelLoader that handles both - but if your workflow was built on the older granular set, this is the loader sitting in it, and it still does exactly the right job.

"FL2VA" is worth unpacking once, because it explains why you'd reach for this node at all. It's H3's image/keyframe-to-video path: you give it a first frame (or a first and last frame), and the model fills in a video between them - motion, audio and all. H3's trick is that it's omni-modal, so even a keyframe-to-video run generates its native stereo sound together with the picture, not as a separate audio pass. That's the capability people came for at launch, and it's the one this loader gates.

Inputs

Same direct-loader contract as its sibling, but pinned to one partition:

• model_root - the weights root (models/MiniMax-H3-INT8-CONVROT, with the old models/MiniMax-H3 layout still accepted).
• dtype - auto, bfloat16, or float16; auto/bfloat16 recommended, with H3's fp32 layers preserved.
• transformer_path - explicit DiT file, defaulting to MiniMax-H3-FL2VA-int8_convrot.safetensors. No auto-swapping between quantized and BF16 here - say what you mean.

The optional attention_backend (auto / sdpa / sage / ck) is the one knob worth touching. ck is Comfy Kitchen's INT8 attention, the same kernel as --use-ck-attention; sage needs SageAttention installed. Both can be meaningful speed wins on a 33B model.

Output: a single h3_model handle for the sampler.

The honest take

This node exists for backward compatibility. It's correct, it's explicit, and there's nothing wrong with it - but the only real difference between this and RHMiniMaxH3Ref2VAModelLoader is which partition name it defaults to and which directory it prefers to resolve. The modern loader figures the partition out from the filename you pick and adds LoRA support, which this one can't. New workflow? Use the modern loader. Old workflow you need to keep working? This one runs fine.

Install and models

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 get 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, or ModelScope in China). The pack needs ComfyUI 0.27+ and a CUDA PyTorch build. And the standing reminder: the MiniMax H3 Community License excludes the US, EU, UK and South Korea from running the local weights - that's a weights issue, not a node issue, but it's yours to navigate.