MICON-Bench: Benchmarking and Enhancing Multi-Image Context Image Generation in Unified Multimodal Models

Abstract

Recent advancements in Unified Multimodal Models (UMMs) have enabled remarkable image understanding and generation capabilities. However, while models like Gemini-2.5-Flash-Image show emerging abilities to reason over multiple related images, existing benchmarks rarely address the challenges of multi-image context generation, focusing mainly on text-to-image or single-image editing tasks. In this work, we introduce MICON-Bench, a comprehensive benchmark covering six tasks that evaluate cross-image composition, contextual reasoning, and identity preservation. We further propose an MLLM-driven Evaluation-by-Checkpoint framework for automatic verification of semantic and visual consistency, where multimodal large language model (MLLM) serves as a verifier. Additionally, we present Dynamic Attention Rebalancing (DAR), a training-free, plug-and-play mechanism that dynamically adjusts attention during inference to enhance coherence and reduce hallucinations. Extensive experiments on various state-of-the-art open-source models demonstrate both the rigor of MICON-Bench in exposing multi-image reasoning challenges and the efficacy of DAR in improving generation quality and cross-image coherence.

Highlights

• 🔍 Six multi-image tasks covering object composition, spatial reasoning, attribute disentanglement, fg/bg swapping, component transfer, and story generation.
• 🤖 Evaluation-by-Checkpoint uses MLLMs as verifiers to automatically score instruction following, identity preservation, and structural coherence.
• 🔁 Dynamic Attention Rebalancing (DAR) offers a training-free plug-in that dynamically modulates attention weights during inference to reduce hallucination and improve cross-image consistency.
• 📊 Rigorous benchmarking across state-of-the-art UMMs reveals key limitations in multi-image context reasoning and demonstrates the effectiveness of DAR.

Status

• Submitted to CVPR 2026

• Role: Co-first author

Recommended citation: **Hang Liu** (co-first author), et al. "MICON-Bench: Benchmarking and Enhancing Multi-Image Context Image Generation in Unified Multimodal Models." Submitted to CVPR 2026.