CLIPLoader (GGUF)

This is the single-encoder loader in city96's GGUF pack - the one you reach for when a model reads your prompt through one text encoder and you want that encoder to be a small, quantized .gguf file instead of the full-precision original. If your model needs two or three or four encoders, there are dedicated Dual/Triple/Quadruple nodes for that. This one is the workhorse for the simple case.

When you'd actually use it

A lot of the newer models run on a single text encoder. Wan video leans on one big T5-family encoder. LTX-Video and PixArt use a lone T5. Several 2026 models encode with a single general-purpose LLM. In all of these, there's exactly one encoder file to load, and it's often the component that quietly decides whether the whole thing fits on your card - the text encoder stopped being an afterthought around the Flux era and has only gotten heavier since.

The trick GGUF gives you is that the encoder is its own file, loaded and quantized independently of the diffusion model. So you can run a full-precision diffusion model against a Q5 or Q8 encoder, free up a couple of gigs, and keep going. That granularity - the Q4-to-Q8 ladder - is the whole reason GGUF became how most people run large models on modest hardware. Q8 is essentially identical to fp16 at half the size; if it fits, just use it.

How it works

Nothing exotic. The node reads a weight file, figures out which quantization level it is, and dequantizes on the fly during inference to hand ComfyUI a normal CLIP object. The one thing worth internalizing: it happily loads both .gguf and regular .safetensors/.bin. So you can point it at a quantized encoder today and a full one tomorrow without changing anything else in your graph. There's a small dequantization overhead versus a plain fp8 file, but on an encoder that's rarely what bottlenecks you.

The inputs and output

• clip_name - the encoder file itself, read from your models/clip folder (newer ComfyUI: models/text_encoders). This is the one you'll be swapping.
• type - the model family, and there's a long list: stable_diffusion, sd3, flux, ltxv, wan, pixart, cosmos, lumina2, hidream, chroma, and more. This is not decoration - it selects the tokenizer and config, so setting it wrong is the fastest way to get output that looks like the model forgot how to read. Match it to whatever you're generating with.

The single output is CLIP, which plugs straight into your CLIPTextEncode node to turn your prompt into conditioning.

Installing it

It ships in city96's ComfyUI-GGUF pack. Via ComfyUI Manager, search "ComfyUI-GGUF" and install. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/city96/ComfyUI-GGUF, then pip install --upgrade gguf, then restart. On Windows portable, clone into ComfyUI/custom_nodes/ComfyUI-GGUF and run the bundled Python against the pack's requirements.txt. That's the only real dependency - the gguf library.

Drop your quantized encoder files in models/clip. city96 hosts a pre-quantized T5-XXL on HuggingFace (t5-v1_1-xxl-encoder-gguf) that covers a lot of the common cases. The node lands under the bootleg category in the node menu.

Common snags

The big one is picking the wrong type - it fails quietly, giving you muddy or nonsensical results rather than an error, so it's the first thing to check when a GGUF encoder "doesn't work." The second is folder confusion: recent ComfyUI moved encoders to models/text_encoders, and if the dropdown is empty, your file is probably in the wrong place or ComfyUI needs a refresh. Finally, don't over-quantize the encoder just because you can - reports on how much a squeezed encoder costs range from "totally fine" to "visibly worse," so treat an aggressive encoder quant as something to A/B, not a free lunch. If fp8 of the same encoder fits, it's simpler and slightly faster; GGUF earns its keep when fp8 won't fit at all.