论文检索

Yufei Gao, Zhengong Cai, Bowei Yang

As microservice architectures become increasingly complex and system events become more frequent, Root Cause Analysis (RCA) has emerged as a critical task to ensure system reliability. However, existing deep learning-based methods often struggle with limited flexibility and a lack of interpretability when addressing complex system failures. Recent efforts to integrate large language models (LLMs) have shown promise in enhancing diagnostic transparency and reasoning capability. However, expansive search spaces, intricate workflows, and entangled constraints constrain practical adoption. We propose RCAFlow, a multi-agent framework that integrates structured workflow knowledge with hierarchical planning to address these challenges. RCAFlow transforms semi-structured documents into behavior tree-style workflows to support interpretable plan generation, employs a Git-inspired branching mechanism for modular and hierarchical task execution with path isolation, and leverages state-aware task execution with semantic analysis to improve result understanding and feedback. We evaluate RCAFlow on three benchmark datasets provided by OpenRCA. Experimental results demonstrate that RCAFlow consistently outperforms existing methods across all datasets. Further ablation studies confirm the effectiveness of each core module, highlighting the reliability, extensibility, and interpretability of RCAFlow to support complex RCA tasks within intelligent IT operations.

Yijia Fan, Jusheng Zhang, Kaitong Cai, Jing Yang, Jian Wang, Keze Wang

Despite recent advancements in 3D-text cross-modal alignment, existing state-of-the-art methods still struggle to align fine-grained textual semantics with detailed geometric structures, and their alignment performance degrades significantly when scaling to large-scale 3D databases. To overcome this limitation, we introduce 3DAlign-DAER, a unified framework designed to align text and 3D geometry via the proposed dynamic attention policy and the efficient retrieval strategy, capturing subtle correspondences for diverse cross-modal retrieval and classification tasks. Specifically, during the training, our proposed dynamic attention policy (DAP) employs the Hierarchical Attention Fusion (HAF) module to represent the alignment as learnable fine-grained token-to-point attentions. To optimize these attentions across different tasks and geometric hierarchies, our DAP further exploits the Monte Carlo tree search to dynamically calibrate HAF attention weights via a hybrid reward signal and further enhances the alignment between textual descriptions and local 3D geometry. During the inference, our 3DAlign-DAER introduces an Efficient Retrieval Strategy (ERS) to leverage efficient hierarchical searching in the large-scale embedding spaces, outperforming traditional methods (eg, KNN) in accuracy and efficiency. Furthermore, to facilitate text-3D alignment research and train our 3DAlign-DAER, we construct Align3D-2M, a large-scale dataset featuring 2M text-3D pairs, to provide sufficient fine-grained cross-modal annotations. Extensive and comprehensive experiments demonstrate the superior performance of our 3DAlign-DAER on diverse benchmarks.

Ke Ding, Brian Parker, Jiayu Wen

Pre-training large language models on genomic sequences has become a powerful approach for learning biologically meaningful representations. While masked language modeling (MLM)-based approaches, such as DNABERT and Nucleotide Transformer (NT), achieve strong performance, they are hindered by inefficiencies due to partial token supervision, pre-training/fine-tuning mismatches, and high computational costs. We introduce NucEL, the first ELECTRA-style pre-training framework for genomic foundation models, which overcomes these challenges. Through a discriminator network identifying tokens modified by a generator, NucEL achieves comprehensive token-level supervision across all sequence positions, thereby markedly improving training efficiency relative to the partial supervision of masked positions inherent in MLM frameworks. By integrating ModernBERT’s architectural advancements, including hybrid local-global attention and flash attention mechanisms, NucEL establishes an optimized BERT architecture for genomic sequence modeling. Unlike traditional methods that tokenize genomic sequences into 6-mers, NucEL implements single-nucleotide tokenization, enabling fine-grained resolution and improving both efficiency and interpretability. Pre-trained on the human genome only, NucEL achieves state-of-the-art performance on benchmark datasets across diverse downstream tasks in both human and non-human species, including regulatory element identification (e.g., promoters, enhancers), transcription factor binding prediction in human and mouse, open chromatin region classification, and histone modification profiles, surpassing MLM-based models of similar size and rivaling models 25 times larger, such as NT. Ablation studies provide critical insights into tokenization and masking strategies, optimizing ELECTRA-style pretraining for DNA sequences. Attention analyses reveal NucEL’s superior ability to capture biologically relevant sequence motifs compared to NT, offering valuable insights into its hierarchical learning process and regulatory element modeling capabilities. This work highlights the potential of ELECTRA-style pretraining as an efficient and effective strategy for advancing genomic representation learning with broad implications for future genomic research.

Longchao Da, David Isele, Hua Wei, Manish Saroya

Being able to anticipate the motion of surrounding agents is essential for the safe operation of autonomous driving systems in dynamic situations. While various methods have been proposed for trajectory prediction, the current evaluation practices still rely on error-based metrics (e.g., ADE, FDE), which reveal the accuracy from a post-hoc view but ignore the actual effect the predictor brings to the self-driving vehicles (SDVs), especially in complex interactive scenarios: a high-quality predictor not only chases accuracy, but should also captures all possible directions a neighbor agent might move, to support the SDVs' cautious decision-making. Given that the existing metrics hardly account for this standard, in our work, we propose a comprehensive pipeline that adaptively evaluates the predictor's performance by two dimensions: accuracy and diversity. Based on the criticality of the driving scenario, these two dimensions are dynamically combined and result in a final score for the predictor's performance. Extensive experiments on a closed-loop benchmark using a real-world dataset show that our pipeline yields a more reasonable evaluation than traditional metrics by better reflecting the correlation of the predictors' evaluation with the autonomous vehicles' driving performance. This evaluation pipeline shows a robust way to select a predictor that potentially contributes most to the SDV's driving performance.

Yujie Chen, Tengfei Ma, Yuansheng Liu, Leyi Wei, Shu Wu, Dongsheng Cao, Yiping Liu, Xiangxiang Zeng

Identifying suitable reaction conditions is critical for chemical synthesis, as they directly affect yield, selectivity, and transformation feasibility. While recent methods have shown promising results, most approaches either encode reactants and products independently or rely on rule-based reaction graphs, both of which constrain the ability of the model to capture condition-relevant structural transformations. In this work, we propose TRACE, a transformation-aware graph refinement framework for reaction condition prediction. TRACE constructs atom-level joint graphs that integrate both reactant and product structures to represent condition-relevant transformations. A structure-aware encoder enriches atom features with local chemical context, followed by a dynamic interaction refinement module that adaptively infers task-specific edges. To further guide the model toward condition-relevant patterns, a mechanism regularized graph encoder incorporates reaction center information, enabling more accurate modeling of transformation mechanisms. Experiments on benchmark datasets show that TRACE achieves state-of-the-art performance across multiple condition types. The integration of transformation-aware refinement leads to improvements in prediction accuracy and generalization, while maintaining robust performance in challenging and realistic synthesis planning scenarios.

Junyang Chen, Yueqian Li, Ka Chung Ng, Huan Wang, Liang-Jie Zhang

The rapid proliferation of social media platforms has led to a surge in multimodal fake news, where deceptive content often combines text and images to mislead audiences. Traditional unimodal detection methods struggle to address the complexity of such content, necessitating holistic multimodal approaches. While the latest advancements in Multimodal Large Language Models (MLLMs) offer new opportunities for enhancing detection performance by analyzing multi-dimensional features, including source credibility, cross-modal contradictions, emotional bias, and manipulative writing patterns, these methods suffer from a key flaw: a susceptibility to hallucinations or erroneous reasoning, which can lead to flawed conclusions and ultimately biased detection results. We propose the Multimodal Fake News Detection via Multi-perspective Rationale Generation and Verification (MMRGV) model to mitigate this challenge. Our method employs a cross-verification mechanism to screen and reconcile contradictions among different rationales, thereby preserving the LLM's analytical advantages while mitigating the impact of erroneous reasoning or hallucinations on the final detection. Subsequently, these optimized rationales are fused via an adaptive weighting strategy to output a robust final prediction. Extensive experiments on three benchmark datasets (Twitter, Weibo, and GossipCop) demonstrate the superiority of our method, achieving state-of-the-art accuracy of 0.9972, 0.9663, and 0.8772, respectively, and significantly outperforming existing baselines. These results validate the effectiveness of multi-perspective rationale generation and cross-verification in enhancing multimodal fake news detection, offering a resilient solution to combat misinformation in the era of generative AI.

Jintae Kim, Keunsoo Ko, Chang-Su Kim

Facial image datasets have become essential resources for various face analysis tasks, but their use raises significant privacy concerns. To address this issue, face obfuscation has emerged as a practical approach to hide identity from humans while retaining cues decipherable by machines. However, existing methods often leave exploitable visual traces, making them vulnerable to reconstruction attacks that restore hidden identity. To address this issue, we propose a frequency-domain manipulation framework, called FreM, which adjusts frequency subbands differently to hide identity, retain machine-decipherable cues, and improve robustness against reconstruction attacks. Specifically, the proposed FreM first decomposes a facial image into frequency subbands and applies subband-adaptive modulation that regulates information according to the characteristics of each subband. The modulation parameters are then refined to yield the reliable obfuscated result. Extensive experiments across multiple face analysis benchmarks demonstrate that FreM achieves superior obfuscation quality and strong robustness against reconstruction attacks.

Eunsu Kim, Junyeong Park, Na Min An, Junseong Kim, Hitesh Laxmichand Patel, Jiho Jin, Julia Kruk, Amit Agarwal, Srikant Panda, Fenal Ashokbhai Ilasariya 等

In a globalized world, cultural elements from diverse origins frequently appear together within a single visual scene. We refer to these as culture mixing scenarios, yet how Large Vision-Language Models (LVLMs) perceive them remains underexplored. We investigate culture mixing as a critical challenge for LVLMs and examine how current models behave when cultural items from multiple regions appear together. To systematically analyze these behaviors, we construct CultureMix, a food Visual Question Answering (VQA) benchmark with 23k diffusion-generated, human-verified culture mixing images across four subtasks: (1) food-only, (2) food+food, (3) food+background, and (4) food+food+background. Evaluating 10 LVLMs, we find consistent failures to preserve individual cultural identities in mixed settings. Models show strong background reliance, with accuracy dropping 14% when cultural backgrounds are added to food-only baselines, and they produce inconsistent predictions for identical foods across different contexts. To address these limitations, we explore three robustness strategies. We find supervised fine-tuning using a diverse culture mixing dataset substantially improve model consistency and reduce background sensitivity. We call for increased attention to culture mixing scenarios as a critical step toward developing LVLMs capable of operating reliably in culturally diverse real-world environments.

Wei Chow, Linfeng Li, Lingdong Kong, Zefeng Li, Qi Xu, Hang Song, Tian Ye, Xian Wang, Jinbin Bai, Shilin Xu 等

Recent advances in diffusion models (DMs) have achieved exceptional visual quality in image editing tasks. However, the global denoising dynamics of DMs inherently conflate local editing targets with the full-image context, leading to unintended modifications in non-target regions. In this paper, we shift our attention beyond DMs and turn to Masked Generative Transformers (MGTs) as an alternative approach to tackle this challenge. By predicting multiple masked tokens rather than holistic refinement, MGTs exhibit a localized decoding paradigm that endows them with the inherent capacity to explicitly preserve non-relevant regions during the editing process. Building upon this insight, we introduce the first MGT-based image editing framework, termed EditMGT. We first demonstrate that MGT's cross-attention maps provide informative localization signals for localizing edit-relevant regions and devise a multi-layer attention consolidation scheme that refines these maps to achieve fine-grained and precise localization. On top of these adaptive localization results, we introduce region-hold sampling, which restricts token flipping within low-attention areas to suppress spurious edits, thereby confining modifications to the intended target regions and preserving the integrity of surrounding non-target areas. To train EditMGT, we construct Crisp-2M, a high-resolution (>1024) dataset spanning seven diverse editing categories. Without introducing additional parameters, we adapt a pre-trained text-to-image MGT into an image editing model through attention injection. Extensive experiments across four standard benchmarks demonstrate that, with fewer than 1B parameters, our model achieves state-of-the-art image similarity performance while enabling faster editing. Moreover, it delivers comparable or superior editing quality, with improvements of 3.6% and 17.6% on style change and style transfer tasks, respectively.

Byungjun Kim, Taeksoo Kim, Junyoung Lee, Hanbyul Joo

Recent progress in 3D reconstruction has made it easy to create realistic digital twins from everyday environments. However, current digital twins remain largely static--limited to navigation and view synthesis without embodied interactivity. To bridge this gap, we introduce Dexterous World Model (DWM), an scene-action-conditioned video diffusion model enabling embodied interaction within static 3D scenes. Given a static 3D scene rendering and an egocentric hand motion sequence, DWM generates temporally coherent videos depicting plausible human-scene interactions. Our approach conditions video generation on (1) static scene renderings following a specified camera trajectory to ensure spatial consistency, and (2) egocentric hand mesh renderings that encode both geometry and motion cues in the egocentric view to model action-conditioned dynamics directly. We train our model on a synthetic human-scene interaction dataset and real-world object manipulation dataset, then evaluate it across both synthetic and real-world egocentric benchmarks. Experiments demonstrate that DWM enables realistic, physically grounded interactions, such as grasping, opening, or moving objects, while maintaining camera and scene consistency. This framework establishes the first step toward video diffusion-based interactive digital twins, enabling embodied simulation and 3D scene interactivity from egocentric actions.

Peter Yongho Kim, Juhyeon Park, Jungwoo Park, Jubin Choi, Jungwoo Seo, Jiook Cha, Taesup Moon

Modeling long-range spatiotemporal dynamics in functional Magnetic Resonance Imaging (fMRI) remains a key challenge due to the high dimensionality of the four-dimensional signals. Prior voxel-based models, although demonstrating excellent performance and interpretation capabilities, are constrained by prohibitive memory demands and thus can only capture limited temporal windows. To address this, we propose TABLeT (Two-dimensionally Autoencoded Brain Latent Transformer), a novel approach that tokenizes fMRI volumes using a pre-trained 2D natural image autoencoder. Each 3D fMRI volume is compressed into a compact set of continuous tokens, enabling long-sequence modeling with a simple Transformer encoder with limited VRAM. Across large-scale benchmarks including the UK-Biobank (UKB), Human Connectome Project (HCP), and ADHD-200 datasets, TABLeT outperforms existing models in multiple tasks, while demonstrating substantial gains in computational and memory efficiency over the state-of-the-art voxel-based method given the same input. Furthermore, we develop a self-supervised masked token modeling approach to pre-train TABLeT, which improves the model's performance for various downstream tasks. Our findings suggest a promising approach for scalable and interpretable spatiotemporal modeling of brain activity. Our code is available at https://github.com/beotborry/TABLeT.

Qianfeng Yang, Qiyuan Guan, Xiang Chen, Jiyu Jin, Guiyue Jin, Jiangxin Dong

Despite significant progress has been made in image deraining, we note that most existing methods are often developed for only specific types of rain degradation and fail to generalize across diverse real-world rainy scenes. How to effectively model different rain degradations within a universal framework is important for real-world image deraining. In this paper, we propose UniRain, an effective unified image deraining framework capable of restoring images degraded by rain streak and raindrop under both daytime and nighttime conditions. To better enhance unified model generalization, we construct an intelligent retrieval augmented generation (RAG)-based dataset distillation pipeline that selects high-quality training samples from all public deraining datasets for better mixed training. Furthermore, we incorporate a simple yet effective multi-objective reweighted optimization strategy into the asymmetric mixture-of-experts (MoE) architecture to facilitate consistent performance and improve robustness across diverse scenes. Extensive experiments show that our framework performs favorably against the state-of-the-art models on our proposed benchmarks and multiple public datasets.

Omid Nejatimanzari, Hojat Asgariandehkordi, Taha Koleilat, Yiming Xiao, Hassan Rivaz

Large vision-language models (VLMs) excel on general benchmarks but often lack robustness in medical imaging, where heterogeneous supervision induces cross-dataset interference and sensitivity to data regime (i.e., how the supervisory signals are mixed). In realistic clinical workflows, data and tasks arrive sequentially, so naive continual training further leads to catastrophic forgetting. To address these challenges, we propose MedQwen, a parameter-efficient medical VLM that couples a spectrally routed Mixture-of-Experts (MoE) with a theoretically grounded scaling rule that aligns low-rank updates with a full-rank, fully fine-tuned MoE, without changing the base architecture. Concretely, we initialize each expert from non-overlapping singular value decomposition (SVD) segments of the pretrained weight and introduce a residual compensation and scaling scheme to enable stable expert specialization and consistent routing under distribution shift.Across 23 medical datasets covering visual question answering, report generation, radiology classification, and hallucination mitigation, MedQwen achieves strong, reliable performance: it approaches full fine-tuning on zero-shot classification with 339xfewer trainable parameters, and reduces sequential forgetting to ~5% where strong baselines degrade by >20-50%.

Jingjie Shang, Tengyu Ma, Heng Zhang, Jinyuan Liu, Risheng Liu, Yuan Wang, Xiaochen Bo

Multi-Exposure Fusion (MEF) seeks to generate a single high-quality image from multiple inputs captured at different exposure levels. Despite substantial progress, most existing approaches depend on statistical metrics that poorly reflect human perceptual preferences. Electroencephalography (EEG) provides a direct physiological window into human cognition, yet its use in low-level vision remains limited due to scarce paired data and the absence of bio-signals during inference. We address these challenges through two key contributions. First, we introduce Cog-Expo, the first dataset capturing human cognitive responses to multi-exposure stimuli, establishing a bridge between neuroscience and computational photography. Second, we propose a bi-level coupled learning framework that leverages this cognitive information without requiring it during inference. A Mental Integrated Transformer serves as the Teacher, incorporating cognitive priors to guide visual feature learning, while a lightweight Student is trained to approximate these cues using only image inputs. Through bi-level optimization, the Teacher learns inherently distillable representations, enabling the Student to emulate cognitive guidance efficiently. Extensive experiments confirm that our method achieves state-of-the-art fusion performance and aligns more closely with human perception.

Shreyas Dixit, Ashhar Aziz, Shashwat Bajpai, Vasu Sharma, Aman Chadha, Vinija Jain, Amitava Das

By 2026, up to 90% of online content may be synthetically generated, raising urgent concerns about the proliferation of AI-driven disinformation. In response, policymakers and technology companies are turning to watermarking as a safeguard: California's Bill AB 321 mandates watermarking of AI-generated media, while firms like Meta and Google have begun deploying watermarking systems to mitigate misuse. However, current watermarking methods remain brittle and vulnerable to attack. In this work, we introduce the visual paraphrase attack, a generative method that removes both visible and invisible watermarks from AI-generated images. The attack proceeds in two steps: (1) generating a descriptive caption from the original image, and (2) feeding this caption into a diffusion-based text-to-image model to produce a visually similar, watermark-free image. Our experiments show that this attack reliably removes watermarks while preserving semantic content of the original image, exposing a critical flaw in existing watermarking strategies. To counter this, we propose PECCAVI, the first watermarking technique explicitly designed to resist visual paraphrasing. PECCAVI embeds robust, high-fidelity watermarks (PSNR > 30 dB) in semantically stable image regions termed Non-Melting Points (NMPs) using multi-channel frequency domain encoding and noisy burnishing to obfuscate watermark locations and hinder reverse engineering. The method is model-agnostic and significantly more resilient than current alternatives. We also release the first benchmark dataset for visual paraphrasing attacks and open-source all code and resources, providing a foundation for future research on robust watermarking in the era of generative AI

Rohit Kundu, Vishal Mohanty, Hao Xiong, Shan Jia, Athula Balachandran, Amit K. Roy-Chowdhury

The proliferation of generative AI has led to hyper-realistic synthetic videos, escalating misuse risks and outstripping binary real/fake detectors. We introduce \texttt SAGA (\underline S ource \underline A ttribution of \underline G enerative \underline A I videos), the first comprehensive framework to address the urgent need for AI-generated video source attribution at a large scale. Unlike traditional detection, \texttt SAGA identifies the specific generative model used. It uniquely provides multi-granular attribution across five levels: authenticity, generation task (e.g., T2V/I2V), model version, development team, and the precise generator, offering far richer forensic insights. Our novel video transformer architecture, leveraging features from a robust vision foundation model, effectively captures spatio-temporal artifacts. Critically, we introduce a data-efficient pretrain-and-attribute strategy, enabling \texttt SAGA to achieve state-of-the-art attribution using only 0.5% of source-labeled data per class, matching fully supervised performance. Furthermore, we propose Temporal Attention Signatures (\texttt T-Sig ), a novel interpretability method that visualizes learned temporal differences, offering the first explanation for why different video generators are distinguishable. Extensive experiments on public datasets, including cross-domain scenarios, demonstrate that \texttt SAGA sets a new benchmark for synthetic video provenance, providing crucial, interpretable insights for forensic and regulatory applications.

Haoze Zheng, Zihao Wang, Xianfeng Wu, Yajing Bai, Yexin Liu, Yun Li, Xiaogang Xu, Harry Yang

Single-image relighting is highly under-constrained: small illumination changes can produce large, nonlinear variations in shading, shadows, and specularities, while geometry and materials remain unobserved. Existing diffusion-based approaches either rely on intrinsic- or G-buffer-based pipelines that require dense and fragile supervision, or operate purely in latent space without physical grounding, making fine-grained control of direction, intensity, and color unreliable. We observe that full intrinsic decomposition is unnecessary for accurate relighting. Instead, sparse but physically meaningful cues--indicating where illumination should change and how materials should respond--are sufficient to guide a diffusion model. Based on this insight, we introduce LightCtrl that integrates minimal physical priors at two levels: a few-shot latent proxy encoder that extracts compact material-geometry cues from limited PBR supervision, and a lighting-aware mask that identifies illumination-sensitive regions and steers the denoiser toward shading-relevant pixels. To compensate for scarce PBR data, we refine the proxy branch using a DPO-based objective that aligns predicted cues with perceptually preferred relighting behavior. We further present ScaLight, a large-scale object-level dataset with systematically varied illumination and complete camera-light metadata, enabling physically consistent and controllable training. Across object- and scene-level benchmarks, our method achieves photometrically faithful relighting with accurate continuous control, surpassing prior diffusion- and intrinsic-based baselines, including gains of up to +2.4 dB PSNR and 35% lower RMSE under controlled lighting shifts.

Suchae Jeong, Jaehwi Song, Haeone Lee, Hanna Kim, Jian Kim, Dongjun Lee, Dong Kyu Shin, Changyeon Kim, Dongyoon Hahm, Woogyeol Jin 等

Vision-language models (VLMs) have achieved impressive results on single-view vision tasks, but lack the multi-view spatial reasoning capabilities essential for embodied AI systems to understand 3D environments and manipulate objects across different viewpoints. In this work, we introduce Cross-View Relations (XVR), a large-scale dataset designed to teach VLMs spatial reasoning across multiple views. XVR comprises 100K vision-question-answer samples derived from 18K diverse 3D scenes and 70K robotic manipulation trajectories, spanning three fundamental spatial reasoning tasks: Correspondence (matching objects across views), Verification (validating spatial relationships), and Localization (identifying object positions). VLMs fine-tuned on XVR achieve substantial improvements on established multi-view and robotic spatial reasoning benchmarks (MindCube and RoboSpatial). When integrated as backbones in Vision-Language-Action models, XVR-trained representations improve success rates on RoboCasa. Our results demonstrate that explicit training on cross-view spatial relations significantly enhances multi-view reasoning and transfers effectively to real-world robotic manipulation.

Ying Shen, Jerry Xiong, Tianjiao Yu, Ismini Lourentzou

Recent advances in generative video modeling, driven by large-scale datasets and powerful architectures, have yielded remarkable visual realism. However, emerging evidence suggests that simply scaling data and model size does not endow these systems with an understanding of the underlying physical laws that govern real-world dynamics. Existing approaches often fail to capture or enforce such physical consistency, resulting in unrealistic motion and dynamics.In this work, we investigate whether integrating the inference of latent physical properties directly into the video generation process can equip models with the ability to produce physically plausible videos. To this end, we propose PHANTOM, a Physics-Infused Video Generation model that jointly models the visual content and latent physical dynamics. Conditioned on observed video frames and inferred physical states, PHANTOM jointly predicts latent physical dynamics and generates future video frames.PHANTOM leverages a physics-aware video representation that serves as an abstract yet informative embedding of the underlying physics, facilitating the joint prediction of physical dynamics alongside video content without requiring an explicit specification of a complex set of physical dynamics and properties. By integrating the inference of physical-aware video representation directly into the video generation process, PHANTOM produces video sequences that are both visually realistic and physically consistent.Quantitative and qualitative results on both standard video generation and physics-aware benchmarks demonstrate that PHANTOM not only outperforms existing methods in terms of adherence to physical dynamics but also delivers competitive perceptual fidelity.

Xinyu Wei, Kangrui Cen, Hongyang Wei, Zhen Guo, Bairui Li, Zeqing Wang, Jinrui Zhang, Lei Zhang

In controllable image generation, synthesizing coherent and consistent images from multiple reference inputs, i.e., **Multi-Image Composition** (MICo), remains a challenging problem, partly hindered by the lack of high-quality training data.To bridge this gap, we conduct a systematic study of MICo, categorizing it into 7 representative tasks and curate a large-scale collection of high-quality source images and construct diverse MICo prompts.Leveraging powerful proprietary models, we synthesize a rich amount of balanced composite images, followed by human-in-the-loop filtering and refinement, resulting in **MICo-150K**, a comprehensive dataset for MICo with identity consistency.We further build a Decomposition-and-Recomposition (**De&Re**) subset, where 11K real-world complex images are decomposed into components and recomposed, enabling both real and synthetic compositions.To enable comprehensive evaluation, we construct **MICo-Bench** with 100 cases per task and 300 challenging De&Re cases, and further introduce a new metric, **Weighted-Ref-VIEScore**, specifically tailored for MICo evaluation.Finally, we fine-tune multiple models on**MICo-150K** and evaluate them on **MICo-Bench**. The results show that MICo-150K effectively equips models without MICo capability and further enhances those with existing skills.Notably, **Qwen-MICo**, fine-tuned from Qwen-Image-Edit, matches **Qwen-Image-2509** in 3-image composition while supporting arbitrary multi-image inputs beyond the latter's limitation.Our dataset and benchmark will be valuable resources for advancing MICo research.