沐阳 H3 · 取模型（挂补丁用）

This node is a walking piece of the pack's history, and it's fine to meet it that way. 沐阳 H3 · 取模型（挂补丁用） (H3ModelFromBundle) does exactly what 沐阳 H3 取模型 does - unwraps the MYANG_H3 bundle from the loader and hands you a plain MODEL socket for the attention/VRAM/LoRA patch chain - and the source comments are honest about why there are two: it's kept for "positional compatibility with early workflows." When the loader moved to lazy loading of ref2va vs fl2va, this node got a kind input bolted on purely so old saved graphs wouldn't shift their widget positions.

So if you're starting fresh, reach for 沐阳 H3 取模型 (H3Model). If you're loading an older Myang workflow, this is the node you'll find wired up, and it still works - it reads the model off the bundle, with a fallback path for the pre-lazy-loading bundle shape in case you've got one of those saved.

Inputs and output

• h3 - the MYANG_H3 bundle from 沐阳 H3 加载器.
• kind - ref2va (default) or fl2va. Match it to the task: motion transfer and continuation run ref2va; first/last-frame and text-to-video run fl2va. Grab the wrong one and your patch chain is dressing a model the job never samples.

Output is one model (MODEL). Feed it into LoraLoaderModelOnly, MiniMaxLowVRAMAttention, SageAttn, SolAttn - whatever your graph uses before the sampler.

Install and the usual gotchas

Same pack, same install: ComfyUI Manager (search "ComfyUI-MiniMaxH3-Myang") or git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang into custom_nodes, restart, hard-refresh if the browser UI looks stale. No extra Python deps; you supply the H3 diffusion model, Qwen text encoder, and both VAEs yourself.

Nothing about this node is worth losing sleep over. It's a small compatibility shim in a pack that's otherwise all-in on one loader producing one bundle - pick whichever "取模型" variant your workflow was saved with and move on. Just keep the kind aligned with your task mode, and remember the H3 weights carry a territory-restricted community license (US/EU/UK/South Korea excluded), which is a thing to check before you build a long project on them.