HuggingFace CLIP Loader

HuggingFace CLIP Loader loads a text encoder (CLIP) from the Hugging Face cache - or a local path - and emits a CLIP object ready to feed your positive/negative conditioning. It's the specialized sibling of the pack's HuggingFace Checkpoint Loader, for when you want just the text-encoding half, and it inherits that node's defining quirk: it never downloads anything. The model must already be sitting in your local HF cache (~/.cache/huggingface), or you point it at a local folder.

Its superpower is the type dropdown. ComfyUI's own CLIPLoader has this too, but this node inherits the full list of known CLIP types from ComfyUI's core loader - everything from stable_diffusion and sd3 to mochi, ltxv, wan, pixart, cosmos, hidream, and two dozen more. That matters because text encoders are not interchangeable: an SD3 conditioning won't work with a Wan text encoder, and ComfyUI needs to know the family to wire the CLIP output into the right downstream nodes. A wrong type quietly produces garbage conditioning instead of an error, so pick deliberately.

The inputs

• repo_id - a HF repo ID, absolute path, or ./relative path from the ComfyUI root.
• type - the CLIP family, chosen from ComfyUI's standard list (25 options).
• subfolder (optional) - defaults to text_encoder, which is where most HF repos keep their CLIP weights. Clever default.
• filename (optional) - auto-detects; for sharded models, give the .index.json.

One output: CLIP.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Duanyll/duanyll_nodepack
cd duanyll_nodepack && pip install -r requirements.txt

or ComfyUI Manager → "Duanyll Nodepack" → install → restart. Needs huggingface-hub for cache resolution, which is in the pack's requirements.

Using it without tripping over it

The pattern mirrors the checkpoint loader: run hf download <repo_id> first, then wire the node. One genuine advantage over stock ComfyUI: because it reads the HF cache, you can load a text encoder without moving files into ComfyUI's models/clip folder - a real convenience if you already maintain an HF cache for other tooling. Just be honest about the tradeoff: the first-run error message ("Run hf download first") can feel like a broken node if you missed the cache-only design. And if you need the encoder plus the diffusion model plus a VAE, the pack's HfCheckpointLoader or HfDiffusionModelLoader gives you more in one node - use this one when a workflow needs the text encoder in isolation, or a nonstandard encoder type the combined loaders can't handle.