TripleCLIPLoader (Dynamic VRAM)

Some models need three text encoders at once, and TripleCLIPLoader (Dynamic VRAM) is the GGUF-aware version of ComfyUI's built-in loader for exactly that. It's the one you'd see in SD3-class workflows - SD3 and SD3.5 famously run three encoders (CLIP-L, CLIP-G, and T5-XXL) merged into a single conditioning. For the rare workflow that does that, this node lets you run all three from quantized or mixed files instead of blowing the full fp16 stack.

Honestly? For most people, the dual loader is the one that shows up in real workflows - Flux's T5 + CLIP-L pair covers a lot of ground, and SD3.5's trio is a heavier niche. But if you're doing SD3, SD3.5, or any architecture whose text encoding is a three-way merge, this is the node the stock templates map to, just with GGUF support bolted on.

How it works

Same machinery as the dual variant with a third file slot. The three dropdowns all pull from the combined models/clip + models/text_encoders list, and each accepts GGUF or safetensors. Unlike the dual and single loaders, there's no type dropdown here - the loader defaults to sd3, which makes sense given the architecture it exists for. The Dynamic VRAM wrapper manages all three encoders through ComfyUI's aimdo memory manager, and the pack reports tensor-loading progress on one bar per file.

The inputs that matter

• clip_name1, clip_name2, clip_name3 - your three encoder files. All GGUF, all safetensors, or a mix.

Output is CLIP, feeding your CLIP Text Encode node.

Installing

The pack install applies: ComfyUI-GGUF in the ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/molbal/ComfyUI-GGUF

Restart ComfyUI and pip install --upgrade gguf.

Where people get burned

• It's an SD3-default node, and the default is the whole interface. If your model uses three encoders but a different combination, make sure this is actually the right loader - there's no type selector to correct a mismatch.
• Dynamic VRAM requires a supporting ComfyUI build - launch without --disable-dynamic-vram, or the loader errors.
• Three encoders is three files. Each must exist in the dropdown and match what the checkpoint expects; a missing or wrong encoder produces broken conditioning, not an obvious crash.
• Standard pack ground rules: _K quants are text-encoder-only here, and scaled-FP8 files don't mix with GGUF in these loaders.