HuggingFace Triple CLIP Loader

Stable Diffusion 3 and its family condition on three text encoders: CLIP-L, CLIP-G, and the big T5-XXL. That trio is why SD3 follows prompts about as well as it does - and why loading it is fiddly, because those three files live in one place in ComfyUI's mental model and another in Hugging Face's repo layout. HfTripleClipLoader sidesteps the whole thing: give it a repo, get a CLIP.

It's the middle sibling of the pack's loader family (dual for Flux, quadruple for HiDream, triple here), and it solves the same annoying problem as the rest: files already sitting in your HF cache shouldn't have to be copied into models/text_encoders by hand.

How it works

The defaults are tuned for stabilityai/stable-diffusion-3-medium: all three repo_id fields point at that repo, the subfolder_1..3 defaults are text_encoders, and the filename defaults are clip_l.safetensors, clip_g.safetensors, and t5xxl_fp16.safetensors - which is exactly how SD3's repo lays out its encoders. So the stock case is: leave the defaults, hit run.

Under the hood it's the pack's shared loader engine - resolve the HF cache snapshot (or a local path), auto-detect or use the explicit filename, merge shards if the checkpoint is split, and hand all three state dicts to Comfy's load_text_encoder_state_dicts. No type dropdown here; the encoders are detected from the state dicts themselves. Output is a single CLIP.

Inputs and outputs

• repo_id_1 / 2 / 3 - each accepts an HF repo ID, an absolute path, or a ./relative path from the ComfyUI root. All three must be set.
• subfolder_1..3 and filename_1..3 - optional; the defaults already match the SD3 repo, but override them when you're loading a different repo or a sharded checkpoint (use the .index.json filename for sharded files).

One CLIP out, wired into CLIPTextEncode as usual.

Installing it

It's part of Duanyll Nodepack:

cd ComfyUI/custom_nodes
git clone https://github.com/Duanyll/duanyll_nodepack

or install "Duanyll Nodepack" through ComfyUI Manager and restart.

Where people get burned

Same as its siblings: the node reads your HF cache, it doesn't download. The description claims it "Downloads a model from Hugging Face," but the code raises FileNotFoundError with Run \hf download <repo>` first` if the repo isn't cached. Get SD3 medium cached once:

hf download stabilityai/stable-diffusion-3-medium

A second trap is the T5 half of the trio. T5-XXL is a ~9.5B encoder - most of SD3's RAM/VRAM footprint is that one file. If you're on a tight card, that's where a quantized T5 or the smaller T5 encoder makes the difference, and it's worth knowing the loader isn't the bottleneck. And as with every loader here: ambiguous folders need an explicit filename, or the node refuses to guess.