--- license: mit library_name: open_clip tags: - image-text-retrieval - retrieval - openclip - vit - icar --- # ICAR Weights (OpenCLIP ViT-L-14, LAION-2B) This repository provides ICAR fine-tuned OpenCLIP ViT-L-14 (LAION-2B) weights, including early-exit variants, used in the paper "Image Complexity-Aware Adaptive Retrieval for Efficient Vision-Language Models". Paper: https://arxiv.org/abs/2512.15372 Code: https://github.com/MikelWL/ICAR These weights are intended to be used with the ICAR codebase. The standard retrieval setting reported in the paper evaluates COCO/Flickr test queries against a gallery augmented with 100k LAION-COCO distractors. ## Contents All weights are stored under the `checkpoints/` directory to match the ICAR codebase layout. ### Baselines (full path) - `checkpoints/baseline_coco/latest_checkpoint.pt` - `checkpoints/baseline_flickr/latest_checkpoint.pt` ### Early-exit variants (ICAR) - `checkpoints/icar_coco/layer_8/latest_checkpoint.pt` - `checkpoints/icar_coco/layer_12/latest_checkpoint.pt` - `checkpoints/icar_coco/layer_16/latest_checkpoint.pt` - `checkpoints/icar_coco/layer_20/latest_checkpoint.pt` - `checkpoints/icar_flickr/layer_8/latest_checkpoint.pt` - `checkpoints/icar_flickr/layer_12/latest_checkpoint.pt` - `checkpoints/icar_flickr/layer_16/latest_checkpoint.pt` - `checkpoints/icar_flickr/layer_20/latest_checkpoint.pt` ## Usage Use the ICAR scripts as the primary entrypoints. See the main repo README for reference scripts and required data paths. Example (mixed evaluation): ```bash 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_weights.pt \ --eval-category ``` ## ICC Weights ICC (Image Complexity Classifier) weights are hosted separately due to non-commercial licensing restrictions on the ImageNet-pretrained ConvNeXt-V2 checkpoint. Please download ICC weights from the dedicated ICC repository. ## License The ICAR weights in this repository follow the OpenCLIP license (MIT). See the OpenCLIP repository for details and attribution requirements.