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

Mingrui Wu; Hang Liu; Jiayi Ji; Xiaoshuai Sun; Rongrong Ji

cs.CVFeb 23, 2026

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

Authors: Mingrui Wu, Hang Liu, Jiayi Ji, Xiaoshuai Sun, Rongrong Ji

Organizations: Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China. · Zhongguancun Academy, Beijing, China.100094.

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. Github: https://github.com/Angusliuuu/MICON-Bench.

Figures & tables

Figure 1: Overview of MICON-Bench and Evaluation Pipeline. MICON-Bench is a comprehensive benchmark designed to evaluate multi-image context generation across six diverse tasks: Object Composition, Spatial Composition, Attribute Disentanglement, Component Transfer, FG/BG Composition, and Story Generation. Each task provides multiple reference images and a compositional prompt requiring the model to integrate, reason, or transfer information across images. To assess model performance, we introduce an Evaluation-by-Checkpoint framework driven by an MLLM, which automatically evaluates generated results along key dimensions, producing an interpretable composite score.

Figure 2: Overview of the proposed Dynamic Attention Rebalancing (DAR) mechanism. Given multiple reference images, DAR first samples query tokens and computes attention maps between sampled queries and reference key tokens. It then applies a dynamic weighting factor to rebalance attention responses, reinforcing relevant reference regions ( green boxes ) while suppressing distractions ( red boxes ). The example at the bottom shows how DAR helps compose a new scene by integrating the right Indian man from Ref Image 1 and the left doctor from Ref Image 2, achieving coherent and faithful visual synthesis.

Figure 3: Visualization examples of our method vs. baseline on MICON-Bench.

Model	Object	Spatial	Attribute	Component	FG/BG	Story	Avg. Score
Nano-Banana [ 11 ]	95.60	93.79	92.13	84.23	83.13	82.84	89.25
GPT-Image [ 14 ]	96.45	94.41	93.39	87.69	90.15	85.99	91.51
UNO [ 38 ]	58.40	66.68	65.28	28.84	20.96	39.08	44.76
DreamOmni2 [ 40 ]	88.24	84.76	85.28	59.64	76.16	59.58	75.56
Qwen-Image-Edit-2507 [ 34 ]	96.52	88.80	78.04	42.68	72.08	63.81	72.96
BAGEL [ 7 ]	87.64	89.96	89.84	52.40	64.64	65.09	73.55

Table 1: Performance comparison of different models on our MICON-Bench.

Method	SINGLE		MULTIPLE			SCENE			Average $\uparrow$
Method	Character	Object	Character	Object	Char.+Obj.	Character	Object	Char.+Obj.	Average $\uparrow$
BAGEL	5.713	6.223	3.028	6.904	6.802	4.244	5.157	6.219	5.536
BAGEL + DAR	6.255	6.082	4.135	7.175	6.870	4.775	4.838	6.301	5.804
OmniGen2	8.184	7.328	6.564	7.994	8.039	6.871	7.898	7.378	7.532
OmniGen2 + DAR	8.299	8.191	6.644	8.418	7.860	7.056	7.965	7.713	7.768

Table 2: Performance of our method on the OmniContext benchmark.

Method	Single-Subject					Multi-Subject					Overall $\uparrow$
Method	DPG	ID-Sim	IP-Sim	AES	AVG (Single) $\uparrow$	DPG	ID-Sim	IP-Sim	AES	AVG (Multi) $\uparrow$	Overall $\uparrow$
BAGEL	93.65	3.11	46.91	47.96	47.91	89.60	2.62	32.39	45.87	42.62	45.26
BAGEL + DAR	94.70	3.61	47.01	48.83	48.54	92.13	4.28	33.38	45.86	43.91	46.23
OmniGen2	92.60	4.25	56.16	57.11	52.53	91.55	11.51	38.40	57.56	49.76	51.14
OmniGen2 + DAR	92.98	5.23	57.66	57.08	53.24	92.56	12.15	39.05	57.15	50.23	51.73

Table 3: Performance comparison on the XVerseBench benchmark.

Model	CLIP Score (T2I) $\uparrow$		CLIP Score (I2I) $\uparrow$		DINO v2 (I2I) $\uparrow$		LPIPS (I2I) $\downarrow$
Model	Object	Spatial	Object	Spatial	Object	Spatial	Object	Spatial
Nano-Banana	0.3505	0.3665	0.9327	0.9382	0.8962	0.9066	0.4934	0.4726
GPT-Image	0.3572	0.3742	0.9089	0.9168	0.8663	0.8815	0.6431	0.6330
Qwen-Image-Edit	0.3558	0.3700	0.9208	0.9292	0.8789	0.8923	0.5542	0.5249
UNO	0.3391	0.3566	0.9007	0.9073	0.8301	0.8366	0.6453	0.6486
DreamOmni2	0.3454	0.3647	0.9133	0.9142	0.8719	0.8787	0.6217	0.6244

Table 4: Evaluation on other metrics, including CLIP score (Text-to-Image and Reference Images-to-Image), DINO v2 (Reference Images-to-Image), and LPIPS (Reference Images-to-Image) under Object Composition and Spatial Geometric Constraints tasks.

Model	Ref=2	Ref=3	Ref=4	Ref=5
BAGEL	88.50	84.37	75.11	66.36
OmniGen2	92.18	89.52	74.92	67.00

Table 5: Analysis of the impact of the numbers of reference images on UMMs in the MICON-Bench Object Composition Task.

Figure 4: DAR suppresses noisy, irrelevant attention ( red boxes ) and re-focuses activations on the correct subjects ( green boxes ), resulting in cleaner referencing and more faithful composition. The blue boxes highlight the target subjects.

Figure 5: The data statistics of MICON-Bench.

Task	#Cases	#Images	#2 Ref	#3 Ref
Object	200	482	118	82
Spatial	200	498	102	98
Attribut	100	300	0	100
Component	240	601	119	121
FG/BG	200	400	200	0
Story	103	237	72	31

Table 6: Statistics of MICON-Bench, including the number of cases, images, and the distribution of two- and three-reference-image settings for each task.

Figure 6: Human-annotated answer set for the Story Inference task. Each example pairs the reference image sequence with canonical human-written descriptions of what should happen next, which are used as targets for MLLM-based evaluation.

# Reference Images	BAGEL		Omnigen2
	Baseline	+OURS	Baseline	+OURS
2	57.13	61.70	72.01	74.71
3	69.72	77.01	97.80	101.58

Table 7: Average inference time (seconds) on Object Composition dataset measured on a single NVIDIA A800-40GB GPU.

$\gamma$	Object	Spatial	Attribute	Component	FG/BG	Story
0.05	86.11	91.68	92.24	56.07	59.29	64.75
0.15	88.04	91.88	90.76	56.06	71.24	66.34
0.25	76.58	85.41	75.43	44.36	64.41	62.86
0.35	53.36	45.78	37.84	23.76	65.14	58.96
0.55	8.39	18.30	0.67	8.39	22.77	40.44

Table 8: Ablation study on the effect of different weight factor

\gamma

Model	Object	Spatial	Attribute	Component	FG/BG	Story	Avg. Score
BAGEL(Generic)	87.72	84.49	87.51	47.33	51.78	63.66	68.52
BAGEL(Specific)	88.17	84.09	85.20	46.09	51.78	63.51	68.01
BAGEL+ DAR (Generic)	91.22	87.81	89.24	48.49	78.22	68.86	75.84
BAGEL+ DAR (Specific)	91.72	87.53	89.42	47.67	80.00	67.51	75.92

Table 9: Ablation study on prompt sensitivity check for the Evaluation-by-Checkpoint.

Model	Object	Spatial	Attribute	Component	FG/BG	Story	Avg. Score
Nano-Banana [ 11 ]	97.88	98.18	72.03	79.41	83.71	79.06	86.63
GPT-Image [ 14 ]	98.85	97.25	83.70	85.26	91.08	81.58	90.77
UNO [ 38 ]	62.30	57.75	55.11	27.56	28.47	38.17	43.87
DreamOmni2 [ 40 ]	88.95	81.69	62.64	65.52	81.01	59.94	75.26
Qwen-Image-Edit-2507 [ 34 ]	98.03	97.57	65.05	57.12	74.95	60.74	77.26
BAGEL [ 7 ]	89.08	89.32	69.05	59.57	72.39	54.32	73.78

Table 10: Performance comparison of different models on our MICON-Bench evaluated by InternVL3.5-38B.

Model	Object	Spatial	Attribute	Component	FG/BG	Story	Avg. Score
Nano-Banana-Pro [ 11 ]	90.00	93.20	92.50	85.31	95.29	87.33	90.61
Nano-Banana [ 11 ]	92.45	90.34	89.23	76.35	87.67	75.00	85.17
GPT-Image [ 14 ]	92.90	93.46	91.73	79.80	87.12	90.00	89.17
UNO [ 38 ]	56.03	54.12	60.10	15.73	25.00	40.09	41.84
DreamOmni2 [ 40 ]	90.00	81.25	83.30	51.98	78.20	60.35	74.18
Qwen-Image-Edit-2507 [ 34 ]	89.40	84.30	75.00	44.76	77.95	59.50	71.82

Table 11: Human Evaluation results on the uniformly sampled subset of the MICON-Bench.

Target Model	Human	Automatic Verifiers
Target Model	Score	Qwen3-VL	InternVL3.5
Nano-Banana-pro	90.61	92.42	90.93
GPT-Image	89.17	89.15	89.05
Nano-Banana	85.17	85.68	85.72
BAGEL + DAR (Ours)	74.08	75.62	75.57
DreamOmni2	74.18	74.58	73.36

Table 12: Reliability Analysis

Task Category	Configuration	Prompt Template
Object Composition	Reference Image	“A photo of {personalized_obj} in {scene}.”
	Task (2 Refs)	“Generate an image that contains both the complete {obj_a} and {obj_b} together in {chosen_scene}.”
	Task (3 Refs)	“Generate an image that contains all three objects ({obj_a}, {obj_b}, and {obj_c}) together in {chosen_scene}.”
Spatial Composition	Reference Image	“A photo of {personalized_obj} in {scene}.”
	Task (2 Refs)	“Generate an image that contains both the complete {obj_a} and {obj_b} together in {chosen_scene}, with {obj_a} positioned to the {spatial_relation} of {obj_b}.”
	Task (3 Refs)	“Generate an image that contains all three objects ({obj_a}, {obj_b}, and {obj_c}) together in {chosen_scene}, with {left_obj} on the left, {center_obj} in the center, and {right_obj} on the right.”

Table 13: Prompt Templates for Benchmark Construction. We detail the reference image prompts and the specific task prompts used for generation, covering object composition, spatial arrangement, attribute disentanglement, component transfer, background replacement, and story inference.

Dimension	ID	Evaluation Question
A. Instruction Following	A_check_1 (H1)	Does the image contain all specified objects as required by the instructions?
	A_check_2	Are the relative arrangement and requested relations between objects correctly followed?
	A_check_3	Are there no obviously extra or missing salient elements?
B. Identity / Fidelity	B_check_1 (H2)	Does each object’s identity strictly match its reference (e.g., category, instance)?
	B_check_2	Are the key attributes (e.g., color, texture, shape) of each object well preserved?
	B_check_3	Are the object details accurate and easily recognizable?

Table 14: Example of Evaluation Checkpoints. Detailed checklist for Object Composition task.

Figure 7: The evaluation prompt used for MLLM scoring.

Figure 8: More visualization examples of our method vs. baseline on MICON-Bench.

Figure 9: More visualization examples of our method vs. baseline on MICON-Bench.

Image Generation

Image generation studies models that create or edit images from text, examples, layouts, and other structured controls.

See all 463 papers

UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models
Feb 9, 2026Cheng Yang, Chufan Shi, Bo Shui +7Unified Multimodal Models Image Generation
University of California San Diego · University of Southern California · University of Illinois Urbana-Champaign +1
Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards
Jun 25, 2026Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar +5Unified Multimodal Models Multimodal Generation
1Mohamed bin Zayed University of Artificial Intelligence · 4Linköping University · 3Australian National University +1
Multimodal Large Language Models for Multi-Subject In-Context Image Generation
Apr 8, 2026Yucheng Zhou, Dubing Chen, Huan Zheng +1Multi-Subject Image Generation Image Generation
SKL-IOTSC, CIS, University of Macau

Feb 9, 2026J/K move · Enter open · S save

UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models

Cheng Yang, Chufan Shi, Bo Shui, Yaokang Wu+6

University of California San Diego · University of Southern California · University of Illinois Urbana-Champaign+1

Unified Multimodal Models Image Generation Multimodal Reasoning Multimodal Foundation Models Multimodal Model Evaluation Multimodal Understanding

Open paper AskPDF arXiv

Unified multimodal models (UMMs) aim to integrate multimodal understanding and generation within a unified architecture, yet it remains unclear to what extent textual and visual modalities are aligned. To investigate this question, we use reasoning-guided image generation as a diagnostic task, where models produce textual reasoning first and then generate images. We introduce UReason, a benchmark for evaluating reasoning-to-generation alignment in this paradigm, consisting of 2,000 human-curated and human-verified instances spanning five reasoning-intensive tasks: Code, Arithmetic, Spatial, Attribute, and Text. To enable controlled analysis, we develop an evaluation framework that compares direct generation, reasoning-guided generation, and decontextualized generation, which conditions only on the refined prompt extracted from reasoning. Across eight widely used open-source UMMs, while we find that reasoning-guided generation yields improvements over direct generation, somewhat surprisingly, decontextualized generation consistently outperforms reasoning-guided generation by a large margin. Our further analyses suggest that the intended visual semantics in textual reasoning are not reliably reflected in the generated images, despite their unified design and training. Overall, UReason serves as a practical litmus test for reasoning-to-generation alignment and provides a challenging benchmark for developing next-generation, more tightly aligned UMMs.

Abstract

Image Generation

Explore similar work

UReason: Benchmarking Reasoning-to-Generation Alignment in Unified Multimodal Models

Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards

Multimodal Large Language Models for Multi-Subject In-Context Image Generation