C$^2$-Routing: Complementary Cropping and Compression for Efficient Vision-Language Inference
C²-Routing adaptively crops or compresses images before vision-language inference, with a fallback gate for risky reductions. Its reported savings are measured against already token-capped inputs across three models and 18 datasets.
C²-Routing chooses between cropping for local detail and compression for broader context, with a safety gate that can fall back to the original input. Across **three VLMs** and **18 datasets**, it reduced FLOPs by 36.53% and visual tokens by 42.36% on average, with a 0.77% mean relative task-score decrease.
For agents that inspect screenshots, documents, or other images, adaptive preprocessing could reduce inference work while preserving relevant detail. The safety gate is part of the approach, so a simple fixed token cap may not capture the same tradeoff.
C²-Routing chooses between cropping for local detail and compression for broader context, with a safety gate that can fall back to the original input. Across **three VLMs** and **18 datasets**, it reduced FLOPs by 36.53% and visual tokens by 42.36% on average, with a 0.77% mean relative task-score decrease. For agents that inspect screenshots, documents, or other images, adaptive preprocessing could reduce inference work while preserving relevant detail. The safety gate is part of the approach, so a simple fixed token cap may not capture the same tradeoff. The comparisons used a full-image baseline capped at 1,024 visual tokens, and experiments ran on RTX 3090 GPUs. The material does not establish how the method performs on other workloads or hardware.
The results make adaptive image preprocessing a measurable option for reducing visual inference work, with the reported average task-score decrease and a fallback gate defining the tradeoff. They support testing the approach on screenshot or document workflows, but do not establish transfer beyond the studied models, datasets, token cap, and hardware.