DiCoBench: Benchmarking Multi-Image Fine-Grained Perception via Differential and Commonality Visual Cues

2026-06-25 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

DiCoBench is a new, comprehensive, multi-image high-resolution benchmark designed to evaluate Multimodal Large Language Models' (MLLMs) ability to autonomously perceive implicit visual cues. Addressing limitations of existing benchmarks that rely on explicit textual cues or low-resolution inputs, DiCoBench features 765 meticulously curated samples. These samples are categorized into two progressive tracks, Differential Visual Cues and Commonality Visual Cues, covering 8 distinct perception tasks. Formulated as a multiple-choice question task and utilizing imagery approaching 2K resolution, it eliminates evaluation metric bias. Extensive evaluation of 18 diverse MLLMs revealed a striking performance gap compared to human accuracy of 98.3%, indicating current top-performing models struggle significantly with micro-scale detail capture. This benchmark aims to drive future research in autonomous, high-resolution multi-image perception.

Key takeaway

For AI Scientists and Machine Learning Engineers developing MLLMs, DiCoBench reveals a significant gap in fine-grained perception, especially with micro-scale details. You should prioritize research into models capable of autonomously perceiving implicit visual cues across high-resolution, multi-image inputs. Your development efforts must focus on improving accuracy beyond the current 98.3% human benchmark, specifically addressing the challenges posed by differential and commonality visual cues.

Key insights

MLLMs significantly underperform humans in high-resolution, multi-image fine-grained perception, highlighting a critical research gap.

Principles

• Fine-grained perception requires implicit visual cues.
• High-resolution inputs are essential for evaluation.
• Multi-image contexts reveal MLLM limitations.

Method

DiCoBench formulates multi-image fine-grained perception as a multiple-choice question task using high-resolution (approaching 2K) imagery. It categorizes 765 samples into Differential and Commonality Visual Cues across 8 tasks.

In practice

• Use DiCoBench to test MLLM fine-grained perception.
• Focus MLLM development on micro-scale detail capture.
• Design MLLMs for implicit visual cue processing.

Topics

• Multimodal Large Language Models
• Fine-Grained Perception
• Visual Benchmarking
• High-Resolution Imagery
• Computer Vision
• Differential Visual Cues
• Commonality Visual Cues

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• LOCUS: Local Visual Cue Search for Enhancing Fine-Grained Perception in Multimodal Large Language Models — LOCUS enhances MLLM fine-grained perception by internalizing local visual cue search via a training-time proxy task.
• WorldBench: A Challenging and Visually Diverse Multimodal Reasoning Benchmark — Visual diversity is crucial for robust multimodal reasoning benchmark development.
• PerceptionDLM: Parallel Region Perception with Multimodal Diffusion Language Models — Multimodal diffusion language models enable efficient parallel visual perception, overcoming autoregressive MLLM limitations.
• VCIFBench: Evaluating Complex Instruction Following for Video Understanding — MLLMs struggle with complex, constraint-rich video instruction following, highlighting a critical evaluation gap.
• MCBench: A Multicontext Safety Assessment Benchmark for Omni Large Language Models — Current Omni LLMs lack robust cross-modal reasoning for safety, struggling to integrate diverse sensory cues effectively.
• Why Vision Language Models Ignore What They See [Munawar Hayat] - 758 — Current VLMs struggle with physics-based generation, visual grounding, and multi-person image fidelity, necessitating improved training and architectural approaches.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.