IP-Comp Load Concept Subspace

If IP-Comp Concept Subspace is the "teach CLIP a concept" node, this one is the "steal someone else's work" node - in the nicest possible way. It loads a concept subspace straight off disk as a .safetensors file, so you never have to write descriptions or run SVD yourself. Want the fur concept? Pick it from a dropdown and move on with your life.

This is the fast path through the whole IP-Comp system, and for most people it's the one that makes the pack usable. The paper's authors ship a whole collection of prebuilt subspaces, and they're all drop-in.

Where the concepts come from

The pack expects subspaces in ComfyUI/models/ip_concept_subspaces/. The official bundle lives on HuggingFace at ReadyON/IP-Composer-Concept-Subspaces, and it's genuinely useful - it ships concepts like fur, age, dog, emotion, flower, material, outfit, outfit_color, person_pose, season, time_of_day, vehicle, weather, and more. That's a lot of composition for zero effort.

To install them:

# inside your ComfyUI folder
mkdir -p models/ip_concept_subspaces
# drop the .safetensors files from ReadyON/IP-Composer-Concept-Subspaces in there

Subspaces you saved yourself with IP-Comp Save Concept Subspace land in the same folder, so this node is also how you reuse your own concepts across projects.

How it works

Under the hood it's a straightforward safetensors load. The node lists every .safetensors file in models/ip_concept_subspaces/, reads the one you pick (including its metadata - rank and which CLIP model built it), and hands back a CONCEPT_SUBSPACE object. That's the exact same type IP-Comp Concept Subspace produces, so the node feeding it doesn't care which path you took.

Inputs: one dropdown, concept_name, listing the files available. Output: one concept_subspace, which wires into IP-Comp Concept Merge. That's the whole graph.

Installing it

The pack itself installs the usual way:

cd ComfyUI/custom_nodes
git clone https://github.com/godmt/ComfyUI-IP-Composer

restart, let ComfyUI Manager pull in open-clip-torch, then drop the subspace files into models/ip_concept_subspaces/ as above. On this path you never touch the CLIP vision loader - the prebuilt subspace already encodes everything you need.

Where people get burned

Empty dropdown. If concept_name shows nothing, the files aren't in models/ip_concept_subspaces/, or ComfyUI loaded before you put them there. Drop the files, then refresh the node list (or restart ComfyUI). The dropdown is built once at startup, not live.
Model mismatch, again. The subspace was built with a specific CLIP vision model, and that choice is baked into its metadata. Feed its output into a merge whose embeddings come from a different vision model and the concept directions point at the wrong space. If you're using the HF bundle, that's ViT-H-14.
Don't overwrite the good ones. The bundle's names are generic (fur, weather). If you save your own concept with the same name, it overwrites the original - a good reason to keep your own save names descriptive.

That's the entire job of this node, and it's refreshingly small. Load, merge, done.