IF-Bench: Benchmarking and Enhancing MLLMs for Infrared Images with Generative Visual Prompting

TL;DR

This paper introduces IF-Bench, a comprehensive benchmark for evaluating multimodal large language models on infrared images, and proposes GenViP, a training-free visual prompting method that improves model performance by translating infrared images into RGB counterparts.

Contribution

The paper presents the first high-quality infrared image understanding benchmark and a novel training-free visual prompting technique to enhance MLLMs' infrared comprehension.

Findings

Model scale and architecture significantly influence infrared image understanding.

GenViP consistently improves MLLMs' performance on infrared tasks.

IF-Bench provides a reliable platform for evaluating multimodal models on infrared data.

Abstract

Recent advances in multimodal large language models (MLLMs) have led to impressive progress across various benchmarks. However, their capability in understanding infrared images remains unexplored. To address this gap, we introduce IF-Bench, the first high-quality benchmark designed for evaluating multimodal understanding of infrared images. IF-Bench consists of 499 images sourced from 23 infrared datasets and 680 carefully curated visual question-answer pairs, covering 10 essential dimensions of image understanding. Based on this benchmark, we systematically evaluate over 40 open-source and closed-source MLLMs, employing cyclic evaluation, bilingual assessment, and hybrid judgment strategies to enhance the reliability of the results. Our analysis reveals how model scale, architecture, and inference paradigms affect infrared image comprehension, providing valuable insights for this area. Furthermore, we propose a training-free generative visual prompting (GenViP) method, which leverages advanced image editing models to translate infrared images into semantically and spatially aligned RGB counterparts, thereby mitigating domain distribution shifts. Extensive experiments demonstrate that our method consistently yields significant performance improvements across a wide range of MLLMs. The benchmark and code are available at https://github.com/casiatao/IF-Bench.