HuggingFace Checkpoint Loader

HuggingFace Checkpoint Loader is the pack's take on the classic Load Checkpoint node, with one crucial difference that will trip you up the first time: it does not download anything. Given a Hugging Face repo_id, it resolves the repo to the local Hugging Face cache on your disk (~/.cache/huggingface) and loads the checkpoint from there. If the model isn't cached, it throws a clear error telling you to run hf download <repo_id> first. That's not a bug - it's the design, and it makes the node offline-friendly and predictable, but it means the name oversells what it does out of the box.

The appeal is that the Hugging Face cache is the ecosystem's upstream - base models and research weights ship there first, as git repos with real revision tracking (see the KB's Hugging Face panel). If you already manage models with huggingface-cli or the HF cache, this node lets ComfyUI read straight from that same store instead of maintaining a second models/checkpoints copy. It also handles sharded checkpoints - you point it at the .index.json and it merges all the shards.

The inputs

• repo_id - a HF repo ID like black-forest-labs/FLUX.1-dev, an absolute local path, or a ./relative path from the ComfyUI root. This is the "download first" field.
• subfolder (optional) - where the model lives inside the repo or path.
• filename (optional) - the file to load; auto-detects if blank. For sharded checkpoints, give it the index file (e.g. model.index.json).

Three outputs: MODEL, CLIP, VAE - the full checkpoint, one node.

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. The node needs huggingface-hub (in the pack requirements) for the cache path, plus ComfyUI's own loader machinery.

The gotcha that defines this node

Before your first use: hf download <repo_id> into the default cache, then point the node at it. Actually getting the model there is on you:

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

If you skip that, you'll hit the "Run hf download first" error - the node intentionally never makes a network call, even under HF_HUB_OFFLINE. This is either a feature (deterministic, works fully offline, no surprise multi-GB downloads at workflow runtime) or a frustration (you expected it to fetch). Compared to stock ComfyUI, which downloads into models/checkpoints on demand, this node is for people who already live in the HF cache world. If that's not you, the built-in loaders are the simpler path. If it is, this is a clean bridge with solid shard handling.