🧠 PRO Unet Loader GGUF

The name says "GGUF loader," and the tooltip just says "Modelo GGUF," but look at the code and this is actually a multi-format model loader: it accepts .gguf, .safetensors, .ckpt, .pt, .pth, and .bin, scanning ComfyUI's unet, diffusion_models, and checkpoints folders and putting everything in one dropdown. If you've got a mixture of quantized and full-precision UNet/diffusion-model files scattered across folders, this node's file browser sees them all. Pick a file, get a MODEL out, wire it to a sampler.

The GGUF half is the headline, though, and it's worth understanding why that matters. GGUF is the single-file quantized format that made Flux-class models (12B+ parameters) runnable on 8–12GB cards - Q8 is effectively indistinguishable from fp16 at half the size, Q5 is the last tier before visible loss, and Q4_K_M is the accepted compromise for tight VRAM. In ComfyUI you'd normally reach for city96's ComfyUI-GGUF pack for that. This node uses it under the hood if it's installed.

How it works

When the selected file ends in .gguf, it loads through load_gguf_sd + GGUFModelPatcher if city96's ComfyUI-GGUF is present, or falls back to the gguf Python package's own reader. Non-GGUF extensions are handled with the standard loaders - safetensors via load_file, checkpoints via torch.load. The important practical detail: GGUF support is conditional. If neither ComfyUI-GGUF nor gguf>=0.6.0 is installed, .gguf files raise a clear ImportError.

The inputs that matter

• unet_name - the dropdown of model files found in your folders. That's the whole required input.

Output: model (MODEL) - exactly what a KSampler or advanced sampler consumes.

Installing it

Install the whole pack first:

cd ComfyUI/custom_nodes
git clone https://github.com/cdanielp/COMFYUI_PROMPTMODELS

Then, for actual GGUF support, install one of the backends:

pip install gguf>=0.6.0
# or, the more complete path:
cd ComfyUI/custom_nodes && git clone https://github.com/city96/ComfyUI-GGUF

Restart ComfyUI afterward. Requires ComfyUI >= 0.26.0.

Gotchas

The biggest trap is the silent expectation: if you install the pack but not a GGUF backend, .gguf files will fail at load time with an ImportError, and it's easy to blame the model file. Check the backend first. Second, this is a loader, not a VRAM miracle - a .gguf file still needs the rest of the pipeline (CLIP, VAE) and a sampler that fits your card. And if you only ever load .safetensors, you don't need the GGUF backends at all; this node works fine as a plain multi-folder UNet loader.