Anymatix UNET Loader

If you've loaded a diffusion model in ComfyUI, you know the standard UNETLoader: pick a .safetensors from a dropdown, optionally pick a weight dtype, done. AnymatixUNETLoader is the Anymatix-pack twin of that node, and it changes exactly one thing: the dropdown becomes a string path. Same class parent, same output, same weight-dtype choices - just addressable by URL, so the model doesn't have to exist on disk when the workflow is written.

That single change is the entire story of this pack, and it's worth understanding because every loader in it follows the same pattern. Anymatix is a desktop app for image and video generation that runs its own ComfyUI and addresses models by URL: a card says "this diffusion model," a fetcher node downloads it (resuming interrupted transfers, verifying size, hashing and deduplicating against models it already holds), and hands a loader a local path. A dropdown can only list what's already in your folders; a string can name a file that's still being fetched. When you see one of these in an exported workflow, the fetcher ahead of it is the real brain - this node just resolves the path to a basename and hands off to ComfyUI's own loader code.

The inputs that matter

• unet_name (STRING) - the path to the diffusion model file, straight from a fetcher output or typed by hand. Resolved by basename against your diffusion_models folder.
• weight_dtype - default, or the fp8 options fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2. This is where you trade a little quality for a lot of VRAM headroom. On a 12GB card, loading a big video or image model in fp8 is often the difference between running and OOMing; fp8_e4m3fn is the usual recommendation, with _fast favouring speed and e5m2 for older-weight compatibility. If you're not VRAM-pressed, default is fine.

Output is a MODEL, which you wire into your sampler's model input exactly as you would from the stock loader.

Install

Under the "Anymatix" category in the anymatix-comfy-nodes pack. ComfyUI Manager: search "anymatix-comfy-nodes" and install, then restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Anymatix/anymatix-comfy-nodes

then restart ComfyUI. The app itself isn't required - the README is clear that the pack exists so Anymatix-exported workflows run on stock ComfyUI.

Gotchas

If you wire in a path but no fetcher ran, you'll get a "file not found" - the loader doesn't download anything itself, it only loads. Let the fetcher execute first (this pack marks fetchers as output nodes specifically so they always run, but if you're hand-building, make sure one's actually in the graph). And if the model never appears despite the file existing, the usual culprit is the basename resolution: the file needs to sit under diffusion_models (or whatever your extra_model_paths.yaml maps), because that's where the loader looks once it's stripped your path down to a name.