Vantage GGUF UNET Loader

GGUF is how a 12GB card runs models that have no business fitting there. This node is a drop-in UNet loader for those quantized files, built on city96's ComfyUI-GGUF (the pack vendors that implementation, credits and all). If you have a .gguf diffusion model sitting in ComfyUI/models/diffusion_models (or unet), this loader's dropdown finds it and hands you a standard MODEL to wire into your sampler - no separate ComfyUI-GGUF install needed.

The why is the GGUF quality ladder. Q8 is basically indistinguishable from fp16 at roughly half the size; Q5 is the last tier before visible loss; Q4_K_M is the accepted compromise when VRAM is tight; Q3 and below are the "I have 6GB and no other option" tier. If the file fits and you have the VRAM, Q8 is the boring correct answer. The trade to know: GGUF runs through a dequantization step, so it's a memory-for-time swap, and LoRA patches on quantized weights pay an extra per-layer dequantize-patch-requantize tax - fine, just slower.

What it needs

• unet_name - a dropdown populated from the GGUF files it can see in your model folders. Drop the .gguf file in diffusion_models (or unet), refresh, and it appears here.

One output: MODEL, ready for a KSampler. That's the whole interface - the loader is thin on purpose.

Under the hood it uses the vendored GGUF ops, loads the state dict, detects the model type, and returns a GGUF-aware model patcher so downstream patching (LoRAs, embeddings) works the way the ecosystem expects. It's the same machinery ComfyUI-GGUF users are used to, packaged inside Vantage-Nodes so you don't have to install the upstream pack separately.

The one thing to remember

GGUF files go in the same folders you're used to, not a special "gguf" subfolder - the loader maps the unet_gguf key onto diffusion_models/unet. And since it needs the gguf Python package (the pack's requirements.txt pins gguf>=0.13.0), a fresh install that skips pip install -r requirements.txt will show an empty dropdown or fail to load. The dependency is real, not optional.

Install

Part of Vantage-Nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/vantagewithai/Vantage-Nodes.git
pip install -r requirements.txt

or ComfyUI Manager → search "Vantage-Nodes" → Install → restart.

Common issues

• Dropdown is empty - no .gguf files in diffusion_models/unet, or the gguf dependency didn't install. Confirm pip show gguf.
• "Could not detect model type" - the GGUF's architecture isn't recognized. Check that the file is a diffusion UNet GGUF and not, say, a text-encoder or VAE quant.
• Slow with LoRAs - expected with quantized weights; drop a quant level to free VRAM for the LoRA pass rather than fighting it.