VELVET VICE MiniMax H3 — AUTO Model Loader

Every MiniMax H3 image-to-video workflow starts with the same problem: a diffusion model that's tens of gigabytes and a filename nobody remembers exactly. The Velvet VICE MiniMax H3 UNET Loader exists so you never have to type it. Set unet_name to AUTO and it hunts down the official H3 file - minimax_h3_fl2va_pruned_int8_convrot.safetensors - anywhere under your diffusion_models folder, subfolders included, and loads it. If that exact file isn't there, it falls back to any local MiniMax H3 FL2V-compatible diffusion model instead of throwing a hissy fit.

It's a thin, honest wrapper around ComfyUI's core UNETLoader. The mechanism is simple: it resolves the filename with the pack's own search logic, then calls through to core's loader with whatever weight_dtype you picked. That's it. No API, no key, nothing exotic. The cleverness is all in the AUTO resolution, which is the part that saves you the "no such file" error at 2am.

The two inputs that matter:

• unet_name - defaults to AUTO. AUTO is the whole point; it matches the official filename by basename in any subfolder. You can also pick any diffusion model you have installed from the dropdown if you want something specific.
• weight_dtype - default, or one of the fp8 variants (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2). If you're squeezed on VRAM, fp8 here is the load-time quantization knob. H3's native diffusion file already ships as a pruned INT8-ConvRot safetensor, so "default" isn't the disaster it is on other models - fp8 is your second rung on the fit ladder, GGUF (handled elsewhere in this pack) is the third.

The single output is a standard MODEL socket, so it wires into anything that eats a model: conditioning, samplers, or the pack's own turbo/router nodes.

Where you'll actually meet this node: most of the time you don't. In the full Velvet VICE H3 workflow the System Hub resolves a backend and a Model Router loads it for you behind a preflight barrier. This loader is the primitive underneath - reach for it when you're hand-building your own H3 graph and want the same filename-proof loading without the whole hub ceremony.

Installing and getting the model

Install once for the whole pack - all these nodes ship together:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

Then restart ComfyUI. Manager users can just search "velvet-vice-minimax-h3" instead. There are zero pip dependencies - the package is pure node glue over ComfyUI core, which is refreshing for a 2026 video pack.

The diffusion model itself you have to source separately (the pack is plumbing, not weights). It goes in ComfyUI/models/diffusion_models/ and can sit in a subfolder - that's exactly the case AUTO was built for. The official weights are MiniMax H3 under its community license, which is a 33B omni-modal model that lands around 42GB in full precision and geofences out the US, EU, UK and Korea - worth knowing before you plan a weekend of rendering.

When it bites

The one real failure mode is a model that isn't there: if AUTO can't find any H3-recognizable diffusion file, the node raises instead of silently loading garbage. That's by design - check that the safetensor is actually in diffusion_models (or a subfolder) and that the name wasn't mangled by your downloader. If you get a warning that a selected model "is not recognizable as MiniMax H3," you've pointed it at a non-H3 file; let AUTO do its job instead.