沐阳 H3 取模型

沐阳 H3 取模型 (H3Model) looks almost useless and is secretly one of the most important nodes in the pack. The 沐阳 H3 加载器 hands you an opaque MYANG_H3 bundle - model weights included but not directly touchable. This node's whole job is to unwrap that bundle and hand you a plain MODEL socket, which is what every patch chain in ComfyUI expects to hang its hats on.

Why would you care? Because H3 is a 33B omni-modal model and the stock ComfyUI path runs it as-is. The Myang example workflows demonstrate a whole shelf of patches - SageAttn, MiniMaxLowVRAMAttention, MiniMaxChunkFeedForward, SolAttn, Spectrum, EasyCache - and every single one of those is a node that takes a MODEL input. You can't patch what you can't reach, so the pattern is always: Loader → H3Model → patch chain → sampler.

The two inputs

• h3 - the bundle from 沐阳 H3 加载器.
• kind - ref2va (default) or fl2va. This is the trap. The tooltip says it plainly: it has to match your task mode. Motion transfer and continuation run the ref2va model; first/last-frame and pure text-to-video run fl2va. Grab the wrong one and your carefully patched attention chain is patching a model that never gets sampled, which you'll notice as "why is none of my patching doing anything."

Output is one model (MODEL), ready for a LoraLoaderModelOnly, an attention patch, a low-VRAM wrapper, or whatever your setup needs before it goes into the sampler.

Install and context

Same as the rest of the pack: ComfyUI Manager search "ComfyUI-MiniMaxH3-Myang", or git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang into custom_nodes, restart, hard-refresh. No extra Python dependencies; the models themselves (H3 diffusion, Qwen text encoder, video + audio VAE) you source yourself.

A small piece of advice from running these graph-heavy long-video packs: don't scatter random patch chains everywhere "for safety." Start with the plain two-segment baseline the README insists on, confirm the seam looks right, then add one patch at a time. Every attention or chunking patch changes memory and quality in ways you'll want to isolate. And remember H3's license geofence (US/EU/UK/South Korea excluded from the local weights) - the pack can't fix that, so check the model's terms before you commit to a workflow built on it.