C$^2$-Routing: Complementary Cropping and Compression for Efficient Vision-Language Inference
The authors report 36.53% fewer end-to-end FLOPs, with a mean relative task-score decrease of 0.77%, against baselines already capped at 1,024 visual tokens.
C²-Routing chooses between cropping and whole-image compression based on the image and question. Cropping emphasizes small targets and local details; compression better preserves broader context and distributed evidence. A safety gate predicts correctness under original and reduced inputs to guide fallback. Across three vision-language models and 18 datasets, the method also reduced visual tokens by 42.36% on average.
ArXiv · AI/CL/LG's note
C²-Routing chooses between cropping and whole-image compression based on the image and question. Cropping emphasizes small targets and local details; compression better preserves broader context and distributed evidence. A safety gate predicts correctness under original and reduced inputs to guide fallback. Across three vision-language models and 18 datasets, the method also reduced visual tokens by 42.36% on average.
ArXiv · AI/CL/LG's note
score 4