TripleCLIPLoader (GGUF)

If you're running SD3 or SD3.5, this is your text-encoder loader. Stable Diffusion 3 was the model that went to three encoders at once - CLIP-L, CLIP-G, and T5-XXL - and this node is the GGUF-aware version of ComfyUI's TripleCLIPLoader that loads all three, with the option to run any of them as a quantized .gguf file. It's a niche node, because SD3.5 itself is a niche these days, but if that's the stack you're on, this is the piece you need.

Why three encoders

SD3's architecture (MMDiT) conditions on three separate text encoders: the two CLIP encoders it borrowed from SDXL - CLIP-L and OpenCLIP-G - plus T5-XXL, the big Google language model that gave it real natural-language understanding and better text-in-image. Three encoders means three files to load and combine, which is exactly what this node handles in one place.

The reason you'd want GGUF here is the usual one: T5-XXL is heavy. The two CLIP encoders are small, but the T5 is the memory hog, and loading a quantized version of it frees VRAM the diffusion model can use. Because each encoder is its own file, you can mix precisions freely - full CLIP-L and CLIP-G alongside a Q5 or Q8 GGUF T5-XXL. Q8 is effectively identical to full precision at half the size, so on SD3.5 it's close to a free saving.

The inputs and output

• clip_name1 / clip_name2 / clip_name3 - the three encoder files, pulled from your models/clip folder (newer ComfyUI: models/text_encoders). The conventional order for SD3 is CLIP-L, CLIP-G, and T5-XXL. The loader reads both .gguf and plain .safetensors/.bin, so a quantized T5 sitting next to full-precision CLIPs is fine.

Note there's no type selector on this node - unlike the single and dual loaders, the three-encoder arrangement is specific enough that it doesn't need one. The single output is CLIP, wired into your CLIPTextEncode (SD3) node.

One SD3-specific gotcha worth its own paragraph

SD3.5 has hard token caps that this node won't warn you about: 77 tokens each on CLIP-L and CLIP-G, 256 on T5. Go over and the image doesn't degrade gracefully - it goes muddy. Long-prompt users deliberately truncate the CLIP inputs (the RES4LYF node set is the usual tool for it). This is the single most-cited reason people who like SD3.5's look give up on it, and it's invisible from the model card. It has nothing to do with GGUF, but if you're setting up a Triple loader you're running SD3.5, so consider yourself warned.

Installing it

It's part of city96's ComfyUI-GGUF pack. ComfyUI Manager: search "ComfyUI-GGUF", install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/city96/ComfyUI-GGUF, then pip install --upgrade gguf, then restart. Windows portable users clone into ComfyUI/custom_nodes/ComfyUI-GGUF and run the embedded Python against the pack's requirements.txt. The gguf library is the only dependency.

Put your GGUF encoder files in models/clip. city96 hosts a pre-quantized T5-XXL on HuggingFace (t5-v1_1-xxl-encoder-gguf), which is the one you'll most likely want here; the CLIP encoders you already have from any SDXL setup. The node appears under the bootleg category.

Common issues

Slot order is the practical one - if SD3.5 output looks off, make sure your CLIP-L, CLIP-G, and T5 are in the slots the workflow expects, and that all three dropdowns are actually populated (an empty dropdown usually means the file's in the wrong folder or ComfyUI needs a refresh). Beyond that, the token caps above will bite you far sooner than anything GGUF-related, and the general quant advice applies: squeeze the T5 first, keep the small CLIPs at full precision, and A/B rather than assuming an aggressive encoder quant is free. macOS Sequoia needs torch 2.4.1 (the pack's issue #107).