Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping

摘要

Multimodal large language models (MLLMs) often miss small details and spatial relations in cluttered scenes, leading to errors in fine-grained perceptual grounding. We introduce AttWarp, a lightweight method that allocates more resolution to query-relevant content while compressing less informative areas, all while pre- serving global context. At test time, AttWarp closes a simple self-correction loop: the MLLM first produces cross-modal attention on the original image, which we use to rectilinearly warp the input and re-run the same frozen model, reallocating resolution toward regions it deems important without changing weights or architecture. This attention-guided warping preserves all original image information but redistributes it non-uniformly, so small objects and subtle relationships become easier for the same model to read while the global layout remains intact. Across nine benchmarks (TextVQA, GQA, DocVQA, POPE, MMMU, MIA- Bench, MMVP, RealWorldQA, BLINK) and four MLLMs (LLaVA, Qwen-VL, InternVL, and InstructBLIP), AttWarp consistently improves accuracy, strengthens compositional reasoning, and reduces hallucinations, outperforming four competitive baselines that manipulate raw images at test time. Together, these results show that attention-guided warping prioritizes information relevant to the query while preserving context, and that the same MLLMs perform better when given such warped inputs. The code and demos are available on the project page: https://dwipddalal.github.io/Attwarp/