HuggingFace Dual CLIP Loader

ComfyUI's built-in DualCLIPLoader only looks in models/text_encoders (or models/clip). Fine, except the moment you've pulled a model with the huggingface_hub CLI, the files are already sitting in ~/.cache/huggingface/hub - and you're stuck copying gigabytes into ComfyUI's folders. This node is the answer to that exact chore: it reads a dual text encoder straight out of your Hugging Face cache (or any local path) and hands you a working CLIP object.

It's the loader you reach for with Flux-family models, which need two encoders (CLIP-L plus the big T5-XXL), and it's a small but real quality-of-life win in any workflow that swaps text encoders often.

How it works

Despite the "Downloads a model from Hugging Face" description, this node does not download anything. It resolves the HF cache layout (models--owner--repo snapshot dirs) and loads what's already there - it even says so in its own error message, telling you to run hf download <repo> first if the repo isn't cached. That's a feature: it works offline and never makes a surprise network call mid-workflow.

From there it auto-detects the checkpoint file in the subfolder you point at (preferring a .index.json for sharded checkpoints, then .safetensors, .sft, .bin, .pt), merges shards if the checkpoint is split, and feeds both encoders into Comfy's load_text_encoder_state_dicts. You get back one CLIP that plugs straight into CLIPTextEncode and the model loader - exactly like the core DualCLIPLoader, minus the folder-juggling.

Inputs and outputs that matter

The two you'll actually touch:

• repo_id_1 / repo_id_2 - a Hugging Face repo ID (black-forest-labs/FLUX.1-dev is the default), an absolute path like /models/flux, or a relative path like ./models/... from the ComfyUI root. Both must resolve; empty repo IDs raise a clear error.
• type - pick flux for FLUX.1 (the default) and sdxl/sd3/hidream/hunyuan_* etc. for the others. This sets which CLIPType Comfy uses to merge the encoders, so it's not decorative.

Defaults for the optional subfolder_1/subfolder_2 (text_encoder and text_encoder_2) match the official FLUX.1 repo layout, so for the stock case you literally hit default, default, default. The filename_1/filename_2 fields are blank by default for auto-detect - fill them in when a folder holds several candidates and the node refuses to guess.

Output is a single CLIP (one wire), and that's it.

Installing it

It ships in Duanyll Nodepack, so either route works:

# Via ComfyUI Manager: search "Duanyll Nodepack", install, restart
cd ComfyUI/custom_nodes
git clone https://github.com/Duanyll/duanyll_nodepack
# restart ComfyUI

The pack's requirements.txt pulls in huggingface-hub and safetensors for this loader family - Manager handles it, a manual clone needs pip install -r requirements.txt if anything's missing.

Where people get burned

The one that trips everyone: the node won't fetch the repo for you. If black-forest-labs/FLUX.1-dev isn't in your HF cache, you get a FileNotFoundError telling you exactly that. Fix it once, before you queue the workflow:

hf download black-forest-labs/FLUX.1-dev

Also watch for the ambiguity error: if a subfolder contains more than one .safetensors and you left filename blank, it refuses to pick. Name the file explicitly and move on. Both are the node being honest, not broken.