CCIPModelLoader

CCIPModelLoader is the least interesting node in this pack and the one you'll trip over first. It doesn't compare anything or compute anything. It just hands the other three nodes - CCIPExtractFeature, CCIPDifference, CCIPSame - a CCIP_MODEL object so they know which character-similarity model to run. The whole pack is dead in the water without it.

The thing that burns people: the ComfyUI nodes are local-model only. The pack README tells a rosier story about auto-downloading from Hugging Face if you don't pass a model directory - that's true for the bundled ccip_lib CLI, but the node code explicitly killed the remote option. If the loader finds nothing, it raises No model folder selected; remote option removed. Please choose a local model folder. So your first job is getting the model files onto disk in the right place.

Where the models go

CCIP (Contrastive Anime Character Image Pre-Training, from deepghs's imgutils project) ships as ONNX files. The loader scans ComfyUI/models/ccip/ and builds the model_folder dropdown from every subfolder it finds. Each folder needs three files, per the pack's own structure docs:

• model_feat.onnx - extracts an embedding from a character image
• model_metrics.onnx - turns two embeddings into a difference score
• metrics.json - holds the default threshold for "same character" verdicts

Grab them from the deepghs/ccip_onnx repo on Hugging Face. The pruned caformer-24 model is the default and the one you'll probably want; it scores about 0.92 F1 on the similarity benchmark while the other options range down to the fast little ccip-caformer-5_fp32:

cd ComfyUI/models
mkdir -p ccip/ccip-caformer-24-randaug-pruned
cd ccip/ccip-caformer-24-randaug-pruned
wget https://huggingface.co/deepghs/ccip_onnx/resolve/main/ccip-caformer-24-randaug-pruned/model_feat.onnx
wget https://huggingface.co/deepghs/ccip_onnx/resolve/main/ccip-caformer-24-randaug-pruned/model_metrics.onnx
wget https://huggingface.co/deepghs/ccip_onnx/resolve/main/ccip-caformer-24-randaug-pruned/metrics.json

Drop the files, reload ComfyUI, and the dropdown fills in. If it's still empty, check the exact path - it's models/ccip/, not models/ccip_onnx/ or whatever you typed first.

Inputs, outputs, install

One input, one output. model_folder is an enum populated from your models/ccip/ subfolders, and it's the only required input - there are no optional ones. Output is a single CCIP_MODEL, which you wire straight into any of the other three nodes' model sockets.

Installing the pack is standard custom-node fare:

cd ComfyUI/custom_nodes
git clone https://github.com/spawner1145/comfyui-ccip

…then restart ComfyUI. ComfyUI Manager may have it under "comfyui-ccip" if you'd rather click - niche packs sometimes lag the registry, so the clone route is the reliable one. The pack's requirements.txt wants onnxruntime>=1.15.0, plus huggingface-hub, scikit-learn, and tqdm; ComfyUI already ships numpy and Pillow. If you run GPU, onnxruntime-gpu is the swap you want, though at 384×384 on character crops even CPU inference is fast enough that you may never bother.

One last note: because each model folder bundles its own metrics.json, swapping models changes the default "same" threshold automatically - which is exactly why CCIPSame has that use_default_threshold switch. Load a different folder and the verdicts can shift slightly without you touching a slider.