Research map / Vision-language and multimodal / Multimodal LLMs

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

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 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 MICON-Bench, a 1,043-case benchmark for multi-image context generation scored by an MLLM Evaluation-by-Checkpoint framework, and DAR, a training-free attention rebalancing method that improves BAGEL and OmniGen2.

Problem

Existing image generation benchmarks focus on text-to-image or single-image editing, leaving multi-image context generation (cross-image consistency, identity, reasoning over several references) poorly measured; open UMMs also distribute attention indiscriminately across reference images, causing ha

Approach

A six-task benchmark with MLLM-verified binary checkpoints for scoring, plus Dynamic Attention Rebalancing that samples query tokens, computes attention to reference tokens, and scales reference keys by a weight (1±γ) for high- and low-attention tokens at inference without training.

Key results

• On MICON-Bench, DAR raises BAGEL average from 73.55 to 76.31 and OmniGen2 from 67.83 to 69.21 (Qwen3-VL judge).
• DAR improves OmniGen2 on OmniContext average from 7.532 to 7.768 and BAGEL from 5.536 to 5.804.
• Human evaluation on a 120-case subset shows MLLM scores deviate from human scores by only +0.67 on average, with matching model rankings.

Surprising findings

• Performance drops sharply as the number of reference images grows (BAGEL object score falls from 88.50 at 2 refs to 66.36 at 5 refs).
• Larger attention scaling (gamma 0.25 to 0.55) collapses performance to near zero, so DAR is highly sensitive to gamma.

Limitations

• MLLM verifier relies on its own perception and can hallucinate, biasing scores.
• DAR effectiveness is bounded by the quality of the base model's attention maps; misreads may be amplified.

Experimental setup

Datasets: OmniContext XVerseBench

Benchmarks: MICON-Bench OmniContext XVerseBench

Metrics: checkpoint pass rate score CLIP score DINO v2 similarity LPIPS DPG ID-Sim IP-Sim AES

Baselines: Nano-Banana GPT-Image UNO DreamOmni2 Qwen-Image-Edit-2507 BAGEL OmniGen2

Builds on: OmniContext BAGEL OmniGen2

Closest papers by meaning

• GIR-Bench: Versatile Benchmark for Generating Images with Reasoning (ICLR 2026 · similarity 0.89)
• MILR: Improving Multimodal Image Generation via Test-Time Latent Reasoning (ICLR 2026 · similarity 0.89)
• WEAVE: Unleashing and Benchmarking the In-context Interleaved Comprehension and Generation (CVPR 2026 · similarity 0.89)
• LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.89)
• Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.88)
• ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.88)