GGUF Loader (Info)

If you run Flux, Wan, or anything big on a 8–12 GB card, you almost certainly run it as GGUF. That's the quantized single-file format llama.cpp invented and city96 brought to diffusion in 2024 - Q8 is basically fp16 at half the size, Q5 is the last tier where you can't see the difference, Q4_K_M is the accepted compromise when VRAM is tight. The catch is that stock ComfyUI only surfaces .gguf files in the folder its current model architecture expects, which means a quant file sitting in the wrong place simply never appears in your dropdown.

EasyGGUFLoader ("GGUF Loader (Info)", category EasyLoaders) is the GGUF half of the Easy Loaders pack, and its party trick is a custom scanner: instead of ComfyUI's per-folder filters, it walks unet, diffusion_models, and checkpoints recursively - subfolders and symlinks included - and lists every .gguf it finds. Drop a quant file anywhere in your model tree and it shows up here, no folder juggling.

How it actually loads

The dropdown input gguf_name is built by that scan. On load, a GGUFReader pulls the file's own metadata - general.architecture, general.name, and tensor count - into the model_info string, then the weights load through this pack's own GGUF engine. The README credits that engine to "gguf-node" (the PIG architecture) by calcuis, a real GGUF converter on HuggingFace, not a made-up name. It's an independent path from city96's ComfyUI-GGUF: weights load through ComfyUI's standard diffusion-model loader with custom GGMLOps operations, wrapped in a GGUFModelPatcher that knows how to patch quantized tensors. What that means in practice: LoRAs work, but every LoRA layer pays the dequantize-patch-requantize tax that all GGUF loaders do - drop a quant level if you're VRAM-capped with LoRAs on.

The pack's own supported-architecture list covers flux, sd1, sdxl, sd3, aura, mochi, ltxv, wan, hidream, pixart, cosmos, and more, so you're not stuck on one model family.

The output that trips people up

Three outputs: model (MODEL), filename_str (STRING), model_info (STRING). Notice what's missing - no clip, no vae. GGUF loaders of this shape load the diffusion model only. Your T5/CLIP text encoders and VAE still load the normal way and wire in separately. If you drag in a Flux workflow, swap its checkpoint loader for this node, and hit run, you'll get dangling clip/vae wires and confusion. Keep the encoders, swap just the model path.

filename_str and model_info behave exactly like the sibling EasyCheckpointLoader node: a scrubbed snake_case filename to feed Save Image's filename_prefix, and a metadata string to view in a Show Text node. The same name_mode dropdown (Formatted / Clean / Original) with prefix, postfix, and max_chars apply - same defaults, same 0-means-unlimited logic.

Installing it

Same as every node in the pack - it all comes in one clone:

cd ComfyUI/custom_nodes
git clone https://github.com/RevengerNick/ComfyUI-Easy-Loaders.git

then restart ComfyUI, or search "Easy Loaders" in ComfyUI Manager. No requirements.txt, no pip installs - the only imports are torch, safetensors, and ComfyUI core, which you already have.

Where people get stuck

• "No .gguf files found" in the dropdown - the scanner found nothing in unet, diffusion_models, or checkpoints. Drop your quant file into one of those (subfolders are fine) and refresh the node.
• "Could not detect GGUF model type" on load - the file's architecture isn't in the supported list. If it's a brand-new model family, that's your answer; the list in the pack's version.json is the ground truth.
• The workflow runs but the image looks wrong or nothing saves - check your encoder wiring, not the quant. This node hands you a model; it does not hand you a complete stack.

One honest word of caution: this pack is young (December 2025), single-author, and has essentially zero community footprint yet - which is exactly why its gguf_name dropdown literally starts with "No .gguf files found" until you add one. If all you want is any GGUF loader, city96's ComfyUI-GGUF is the battle-tested default and the safe recommendation. Reach for this one when you want the wider folder scan and the filename-plus-info convenience in the same node, and you're comfortable being an early adopter.