Research map / Vision-language and multimodal / Multimodal LLMs

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

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces IF-Bench, a 499-image, 680-VQA benchmark for infrared image understanding evaluated on 40+ MLLMs, and GenViP, a training-free method that translates infrared inputs into RGB counterparts with an editing model (Qwen-Edit-2509 fine-tuned on 50k RGB-T pairs) to boost MLLM performance.

Problem

Multimodal LLMs are trained mostly on RGB images, and their ability to understand infrared images is unexplored, with domain shift degrading performance on infrared inputs.

Approach

Use a generative image-editing model to translate infrared images into spatially and semantically aligned RGB images, then feed them alongside the original infrared image (plus a textual prior) to any MLLM at inference time, without training the MLLM.

Key results

• Qwen3-VL-235B-A22B-Instruct scores 83.7 on IF-Bench and reaches 84.4 with GenViP, the top score among evaluated models; Doubao-Seed-Vision-1.6-250815 scores 84.2 without GenViP.
• GenViP with Qwen-Edit-2509-FT improves Qwen2.5-VL-7B from 71.1 to 74.2 (+3.1), outperforming Seedream 4.0 (73.8) and Gemini-2.5-Flash (73.6) as editing models.
• Qwen3-VL-8B-Instruct with GenViP reaches 80.3, matching InternVL3-78B and GLM-4.5V-106B-A12B-Thinking without GenViP while using about one-tenth of their parameters.

Surprising findings

• Thinking mode yields only a modest average gain (+0.34) and hurts Qwen3-VL performance, and it reduces accuracy on target localization and action recognition.
• Under a stricter criterion requiring correct answers across all option permutations, average scores drop by up to about 18 points (e.g., Qwen2.5-VL-7B from 71.1 to 53.3).

Limitations

• IF-Bench contains a relatively limited number of images (499) and questions (680).
• Does not cover some more challenging task types.
• GenViP gains diminish as model scale increases, and it requires an extra editing-model call at inference.

Experimental setup

Datasets: InfPre (23 infrared datasets) RGB-T image pairs (curated, 50,000 pairs)

Benchmarks: IF-Bench

Metrics: accuracy average accuracy across 10 dimensions stricter all-permutation accuracy

Baselines: Seedream 4.0 Gemini-2.5-Flash Doubao-Seed-Vision-1.6 Gemini-2.5-Pro Qwen-Edit-2509

Builds on: MM-Bench circular evaluation InfPre Qwen-Edit-2509

Closest papers by meaning

• Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision (ICLR 2024 · similarity 0.89)
• WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation (CVPR 2026 · similarity 0.89)
• GIR-Bench: Versatile Benchmark for Generating Images with Reasoning (ICLR 2026 · similarity 0.88)
• Learning Interleaved Image-Text Comprehension in Vision-Language Large Models (ICLR 2025 · similarity 0.88)
• SEED-Bench: Benchmarking Multimodal Large Language Models (CVPR 2024 · similarity 0.88)
• JourneyBench: A Challenging One-Stop Vision-Language Understanding Benchmark of Generated Images (NeurIPS 2024 · similarity 0.87)