icc-weights

ICC Weights (ConvNeXt V2)

This repository provides the ICC (Image Complexity Classifier) weights used in "Image Complexity-Aware Adaptive Retrieval for Efficient Vision-Language Models".

Paper: https://arxiv.org/abs/2512.15372 Code: https://github.com/MikelWL/ICAR

The ICC is a ConvNeXt V2 classifier fine-tuned from an ImageNet-22K pretrained checkpoint. It is used to route images between early-exit and full-path inference in ICAR.

Contents

This repository ships a single file at the repo root:

• icc.pt

Usage

Point ICAR to the ICC weights with --icc-checkpoint:

python scripts/evaluate_mixed_preprocessed.py \
  --config icar/configs/coco.yaml \
  --checkpoint checkpoints/icar_coco/layer_12/latest_checkpoint.pt \
  --base-dataset mscoco \
  --base-data-root /path/to/coco-images \
  --laion-data-root /path/to/laion_coco_100k \
  --complexity-scores /path/to/laion_coco_100k_metadata/complexity_scores.json \
  --early-exit-layer 12 \
  --use-icc-routing \
  --icc-checkpoint /path/to/icc.pt

License

The ICC weights are derived from ImageNet-pretrained ConvNeXt V2 models, which are licensed under CC-BY-NC. Please use these weights for non-commercial research purposes and provide appropriate attribution.