MICON-Bench: Multi-Image Context Benchmark

• MICON-Bench is a benchmark for multi-image context image generation in UMMs, emphasizing cross-image composition, contextual reasoning, and identity preservation.
• It comprises 1,043 cases across six tasks and employs an MLLM-driven Evaluation-by-Checkpoint framework along with Dynamic Attention Rebalancing (DAR) for enhanced inference.
• The framework systematically probes multi-reference fusion challenges, improving spatial, attribute, and narrative consistency in generated images.

MICON-Bench is a benchmark and evaluation framework for multi-image context image generation in Unified Multimodal Models (UMMs), introduced in “MICON-Bench: Benchmarking and Enhancing Multi-Image Context Image Generation in Unified Multimodal Models” by Mingrui Wu, Hang Liu, Jiayi Ji, Xiaoshuai Sun, and Rongrong Ji, with affiliations including Xiamen University and Zhongguancun Academy (Wu et al., 23 Feb 2026). It is designed to measure and enhance a setting in which a model receives two or more related reference images together with a textual instruction and must synthesize a single image that is semantically aligned and visually consistent with those references. The framework combines a six-task dataset, an MLLM-driven Evaluation-by-Checkpoint verifier, and a training-free inference mechanism called Dynamic Attention Rebalancing (DAR). Its stated focus is on cross-image composition, contextual reasoning, and identity preservation, with the explicit goal of exposing failure modes that are not systematically probed by benchmarks centered on text-to-image generation or single-image editing (Wu et al., 23 Feb 2026).

1. Problem setting and scope

Multi-image context image generation asks a model to integrate and reason across multiple visual references while obeying a textual instruction. The task demands include preserving identities and attributes taken from specific images, composing foregrounds and backgrounds across references, arranging multiple entities under spatial constraints, transferring components and styles, and maintaining cross-reference consistency (Wu et al., 23 Feb 2026).

The paper distinguishes this setting from two adjacent problem classes. Traditional text-to-image generation grounds only on a single text prompt, whereas single-image editing modifies one image conditioned on text or an auxiliary mask. By contrast, multi-image context generation requires a model to fuse and disambiguate information from multiple visual sources, reason about inter-image relationships such as identity, attributes, and geometry, and produce one coherent output that respects constraints drawn from each reference (Wu et al., 23 Feb 2026).

The motivation for a new benchmark is framed against existing benchmarks such as Geneval, T2ICompBench, ImgEdit-Bench, and EMU Edit, which are described as largely targeting text alignment, realism, or single-image faithfulness rather than systematic evaluation of cross-image composition, multi-reference consistency, or contextual narrative reasoning. MICON-Bench is therefore positioned as a direct probe of three core capabilities: whether a model can correctly combine subjects and backgrounds across references without contradictions, whether it can follow spatial relations, causal logic, and story progression implied by multi-image inputs, and whether it can preserve instance-level identity and fine-grained attributes from specific references despite distractors (Wu et al., 23 Feb 2026).

A plausible implication is that the benchmark is not merely an expanded prompt suite but an attempt to isolate failure modes specific to multi-source visual grounding, especially when identity, geometry, and causality must be inferred jointly.

2. Benchmark composition

MICON-Bench contains 1,043 cases and 2,518 images spanning six tasks. The tasks are designed to stress distinct but overlapping forms of multi-image reasoning, from subject-scene composition to part-level transfer and narrative continuation (Wu et al., 23 Feb 2026).

Task	Core objective	Cases / images
Object Composition	Combine one or more subjects with a background scene while preserving identities and salient attributes	200 / 482
Spatial Composition	Arrange multiple objects under explicit geometric relations with identity preservation	200 / 498
Attribute Disentanglement	Recombine subject, style, and background from three distinct references	100 / 300
Component Transfer	Extract specific elements from one reference and apply them to a target subject in another	240 / 601
FG/BG Composition	Extract the foreground from one reference and combine it with the background of another	200 / 400
Story Generation	Infer and render what happens next from two or three reference panels	103 / 237

Each task is defined in terms of a target relationship among references.

Object Composition uses 2 or 3 reference images depicting a subject or a scene and asks for one synthesized image containing all specified subjects in the designated environment. The representative prompt is: “Generate an image that contains both the complete {obj_a} and {obj_b} together in {chosen_scene}.” For three references, the formulation becomes “all three objects ({obj_a}, {obj_b}, and {obj_c}) together in {chosen_scene}.” The relationship of interest is subject-to-scene composition with identity and attribute fidelity (Wu et al., 23 Feb 2026).

Spatial Composition extends composition by imposing explicit geometric constraints such as left, right, or front. The representative prompt is: “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}.” For three references, the prompt specifies a left-center-right arrangement. The targeted relationship is multi-object spatial geometry under physical plausibility (Wu et al., 23 Feb 2026).

Attribute Disentanglement uses three references, labeled A, B, and C, where A provides the subject, B the style, and C the background. The representative prompt is: “Generate an image of the {main_object} from image A, using the visual style from image B, and placing it in the {specific_background} environment from image C.” The task probes cross-image attribute disentanglement and transfer with identity fidelity (Wu et al., 23 Feb 2026).

Component Transfer focuses on part-level control. It asks the model to extract specific elements such as clothing, accessories, or parts from one reference and apply them to a target subject in another. The benchmark defines both a complex mode and a simple mode, with representative prompts including “Extract {elements_desc} from {source_desc} in Image {source_label}, then apply these elements to {target_desc} in Image {target_label}” and “Extract only the {local_element} from the subject in Image A, then apply this element to the subject in Image B.” This targets fine-grained identity and attribute control (Wu et al., 23 Feb 2026).

FG/BG Composition isolates foreground-background separation. It uses two references and asks the model to cleanly extract the foreground from one image and blend it into the unchanged background of the other. The representative prompt is: “Generate an image where you cleanly extract the {obj_a} from image A and replace the {obj_b} in image B. The background from image B should remain unchanged.” The targeted relationship is foreground-background separation with seamless blending (Wu et al., 23 Feb 2026).

Story Generation differs from the other five tasks by requiring causal and temporal inference. It uses two or three reference panels from four-panel comics generated by GPT-4o-Image and Nano-Banana, with human-filtered story texts and DiffBIR enhancement for clarity. The representative prompt is: “Given the reference images, infer and generate a realistic photo of what might happen next.” The relevant relationship is cross-image causal reasoning, temporal consistency, and narrative closure (Wu et al., 23 Feb 2026).

3. Data construction and curation

For the five compositional tasks, the reference images are generated via Qwen-Image from carefully constructed pools of subjects, attributes, spatial relations, and scenes using aligned prompt templates (Wu et al., 23 Feb 2026). For Story Generation, short narratives are LLM-generated and human-filtered; reference panels are then produced by GPT-4o-Image and Nano-Banana, restored with DiffBIR, and filtered by humans for narrative and visual consistency (Wu et al., 23 Feb 2026).

The curation procedure uses a two-stage filtering pipeline that merges MLLM coarse verification with meticulous human review to ensure fidelity and diversity. The paper does not specify train/val/test splits and states that the benchmark is designed for evaluation. It also states that licensing details are not explicitly specified in the paper and should be checked in the repository (Wu et al., 23 Feb 2026).

Per-task reference cardinalities are explicit. Object Composition has 118 cases with 2 references and 82 with 3 references. Spatial Composition has 102 with 2 references and 98 with 3 references. Attribute Disentanglement uses 3 references in all 100 cases. Component Transfer has 119 cases with 2 references and 121 with 3 references. FG/BG Composition uses 2 references in all 200 cases. Story Generation has 72 cases with 2 references and 31 with 3 references (Wu et al., 23 Feb 2026).

These design choices indicate that MICON-Bench is structured around controlled compositional factors rather than unrestricted web-scale collection. This suggests an emphasis on diagnostic benchmarking: the benchmark aims to separate errors in identity binding, attribute transfer, spatial arrangement, and causal continuation rather than to maximize ecological diversity.

4. Evaluation-by-Checkpoint

The evaluation framework is MLLM-driven. For each case, the instructions, the reference images, and the model’s generated image are fed into an MLLM verifier, with Qwen3-VL-32B-Instruct as the default. The verifier checks task-specific binary checkpoints grouped under seven dimensions: Instruction Following, Identity/Fidelity, Structure/Geometry, Cross-Reference Consistency, Causality, Text Grounding, and Overall Usability (Wu et al., 23 Feb 2026).

Each checkpoint yields pass or fail, and the proportion of satisfied checkpoints within a dimension is converted to a score in $[0,1]$. One hard-constraint checkpoint per dimension can cap that dimension’s score at 0.4 if failed. The active dimension scores are then aggregated and linearly rescaled to $[0,100]$ (Wu et al., 23 Feb 2026).

For Story Generation, the benchmark combines checkpoint scoring with matching against a human-authored answer set. The paper gives the formula

$$S_{\text{story}} = 0.4 \cdot S_{\text{checkpoint}} + 0.6 \cdot S_{\text{answer}}.$$

This weighting is intended to ensure both causal progression and narrative plausibility (Wu et al., 23 Feb 2026).

The rubric assigns distinct semantic roles to the dimensions. Identity/Fidelity emphasizes instance-level identity, category correctness, and key attribute preservation such as color, texture, and shape. Structure/Geometry verifies spatial relations and physical plausibility. Cross-Reference Consistency checks contradictions across references. Causality verifies temporal logic. Text Grounding checks textual content. Overall Usability measures naturalness and coherence (Wu et al., 23 Feb 2026).

The paper also notes that checkpoint questions may be phrased generically or specifically, for example “Does the image include all specified objects?” versus “Does the image include a giraffe and a wooden chair?”, and reports stability across phrasing via a prompt sensitivity study. Reliability is evaluated on a sampled 120-case subset using three expert annotators with the same binary hard-constraint rubric. The automated verifier’s average deviation from human scores is +0.67, and model rankings are reported as consistent. Replacing Qwen3-VL with InternVL3.5-38B on the full dataset preserves relative rankings, and five independent runs evaluating BAGEL with Qwen3-VL differ by at most 0.04 (Wu et al., 23 Feb 2026).

A central limitation is also explicitly stated: as with any MLLM-based evaluation, verifier hallucinations can propagate, and the framework’s reliability depends on the MLLM’s perception. The paper argues that hard constraints and binary checkpoints improve robustness, but it does not treat verifier dependence as eliminated (Wu et al., 23 Feb 2026).

5. Dynamic Attention Rebalancing

Dynamic Attention Rebalancing (DAR) is a training-free, plug-and-play mechanism applied during inference. Its purpose is to adjust attention to reference image tokens so that UMMs focus on relevant regions across multiple references. The method identifies over-attended and under-attended areas via sampled attention maps and then reweights keys to amplify semantically relevant tokens and suppress distractors, with the stated effect of reducing cross-image hallucinations and improving identity and spatial coherence (Wu et al., 23 Feb 2026).

The query sampling rule is given as

$$\left\{\left\lfloor i \cdot \frac{L_q - 1}{m - 1} \right\rfloor \right\}_{i=0}^{m-1}.$$

The paper defines token relevance by

$$r_k = \sum_{i=1}^{m} \sum_{h=1}^{H} \tilde{A}_{i,h,k},$$

followed by min-max normalization,

$$\hat{r}_k = \frac{r_k - \min_j r_j}{\max_j r_j - \min_j r_j}.$$

The algorithmic description then thresholds $\hat{r}_k$ with $\tau_{\mathrm{high}}$ and $\tau_{\mathrm{low}}$ to assign weights $w_k \in \{1-\gamma, 1, 1+\gamma\}$, recomputes full attention from $[0,100]$0 to the reweighted keys $[0,100]$1, and continues the generation process (Wu et al., 23 Feb 2026).

The operational steps are explicit:

1. Sample $[0,100]$2 representative query tokens uniformly from $[0,100]$3 using the stated indices.
2. Compute attention from sampled queries to reference keys across $[0,100]$4 heads.
3. Sum and min-max normalize per reference token to obtain $[0,100]$5.
4. Threshold $[0,100]$6 with $[0,100]$7 and $[0,100]$8 to assign weights.
5. Recompute full attention using the reweighted keys and proceed with generation (Wu et al., 23 Feb 2026).

The default hyperparameters are $[0,100]$9, $$S_{\text{story}} = 0.4 \cdot S_{\text{checkpoint}} + 0.6 \cdot S_{\text{answer}}.$$0, $$S_{\text{story}} = 0.4 \cdot S_{\text{checkpoint}} + 0.6 \cdot S_{\text{answer}}.$$1, and $$S_{\text{story}} = 0.4 \cdot S_{\text{checkpoint}} + 0.6 \cdot S_{\text{answer}}.$$2. On Object Composition with A800-40GB GPUs, DAR adds approximately 5–10% runtime, with examples reported as BAGEL 57.13s → 61.70s and OmniGen2 72.01s → 74.71s for 2 references (Wu et al., 23 Feb 2026).

DAR is applied to the open-source UMMs BAGEL and OmniGen2, and is reported to generalize on OmniContext and XVerseBench. The paper characterizes it as conceptually compatible with other UMMs that expose cross-attention over image tokens (Wu et al., 23 Feb 2026).

Ablation results show that performance declines as the number of references increases; one cited example is BAGEL Object task: 2 refs 88.50 → 5 refs 66.36, highlighting the challenge of multi-source fusion. The weight factor $$S_{\text{story}} = 0.4 \cdot S_{\text{checkpoint}} + 0.6 \cdot S_{\text{answer}}.$$3 is reported to provide the best overall trade-off, whereas excessively large $$S_{\text{story}} = 0.4 \cdot S_{\text{checkpoint}} + 0.6 \cdot S_{\text{answer}}.$$4 harms robustness; the paper notes that $$S_{\text{story}} = 0.4 \cdot S_{\text{checkpoint}} + 0.6 \cdot S_{\text{answer}}.$$5 yields very low accuracies. Qualitatively, DAR is reported to suppress spurious attention on irrelevant background regions or subjects and to improve localization of intended reference entities (Wu et al., 23 Feb 2026).

6. Empirical findings, usage, and limitations

The benchmark evaluates several model categories: proprietary or native UMMs (Nano-Banana, GPT-Image), diffusion-based systems (UNO, DreamOmni2, Qwen-Image-Edit-2507), and open-source UMMs (BAGEL, OmniGen2) together with their DAR-augmented variants (Wu et al., 23 Feb 2026).

On MICON-Bench with the Qwen3-VL verifier, GPT-Image achieves the highest reported average score at 91.51, with task scores of Object 96.45, Spatial 94.41, Attribute 93.39, Component 87.69, FG/BG 90.15, Story 85.99. Nano-Banana reports Object 95.60, Spatial 93.79, Attribute 92.13, Component 84.23, FG/BG 83.13, Story 82.84, Avg 89.25. Among diffusion-based systems, UNO reports Avg 44.76, DreamOmni2 75.56, and Qwen-Image-Edit-2507 72.96. Among open-source UMMs, BAGEL improves from 73.55 to 76.31 with DAR, a gain of +2.76, while OmniGen2 improves from 67.83 to 69.21, a gain of +1.38 (Wu et al., 23 Feb 2026).

The reported gains are task-dependent. DAR gains are described as notable in Component and FG/BG tasks for OmniGen2, and in FG/BG and Story for BAGEL. On OmniContext, DAR improves BAGEL average 5.536 → 5.804 and OmniGen2 7.532 → 7.768. On XVerseBench, DAR improves BAGEL overall 45.26 → 46.23 and OmniGen2 51.14 → 51.73, with consistent gains in identity- and attribute-similarity metrics (Wu et al., 23 Feb 2026).

For Object and Spatial tasks, the benchmark also reports auxiliary metrics. CLIP T2I changes from 0.3586/0.3748 → 0.3612/0.3761 for BAGEL with DAR, and from 0.3646/0.3828 → 0.3648/0.3828 for OmniGen2. The paper further states that CLIP I2I, DINOv2 I2I, and LPIPS I2I improve on most measures for BAGEL and OmniGen2, indicating stronger semantic alignment and visual quality (Wu et al., 23 Feb 2026).

The practical workflow is straightforward. The repository provides the dataset, prompt templates, evaluation scripts, and verifier configurations, with Qwen3-VL-32B-Instruct as default and InternVL3.5-38B supported. The stated procedure is: generate images for each benchmark case using the provided prompts and references; run the Evaluation-by-Checkpoint script to feed inputs, references, outputs, and checkpoints to the MLLM verifier; then obtain dimension-wise scores, apply hard-constraint capping and rescaling, and compute final aggregate scores, with Story using the 40%/60% checkpoint/answer-set combination. Experiments were run with 4× NVIDIA A800 (40G), and the paper notes that the MLLM evaluator is stable across runs, with a maximum score discrepancy of 0.04 in repeated trials (Wu et al., 23 Feb 2026).

The limitations are explicit. Evaluation relies on an MLLM’s visual reasoning, so verifier hallucinations may bias scores even with hard constraints, binary checks, and human alignment. The tasks focus on image synthesis from multi-image inputs, leaving domains such as video, 3D, and long-range temporal narrative open. DAR depends on the quality of the base attention maps; if a model misinterprets references or misses fine-grained semantics, reweighting may amplify incorrect signals. The paper identifies stronger calibrated verifiers with counterfactual tests, broader multi-image tasks, identity- and geometry-aware training objectives, attention diagnostics for interpretability, and scaling DAR to varied architectures and diffusion backends as future directions (Wu et al., 23 Feb 2026).

The release is centered on the public GitHub repository at https://github.com/Angusliuuu/MICON-Bench, which provides prompt templates, references, evaluation scripts, and verifier configurations. The paper states that exact usage terms and licensing should be verified in the repository itself (Wu et al., 23 Feb 2026).