Compositional Image Generation

While recent advancements in multimodal language models have enabled image generation from expressive multi-image instructions, existing methods struggle to maintain performance under complex interleaved instructions. This limitation stems from the structural separation of images and text in current paradigms, which forces models to bridge difficult long-range dependencies to match descriptions with visual targets. To address these challenges, we propose \texttt{I}mages i\texttt{N} \texttt{SE}n\texttt{T}ences (\textit{a.k.a}, INSET), a unified generation model that seamlessly embeds images as native vocabulary within textual instructions. By positioning visual features directly at their corresponding semantic slots, INSET leverages the contextual locality of transformers for precise object binding, effectively treating images as dense, expressive language tokens. Furthermore, we introduce a scalable data engine that synthesizes 15M high-quality interleaved samples from standard image and video datasets, utilizing VLMs and LLMs to construct rich, long-horizon sequences. Evaluation results on InterleaveBench demonstrate that INSET significantly outperforms state-of-the-art methods in multi-image consistency and text alignment, with performance gaps widening as input complexity increases. Beyond standard generation, our approach inherently extends to multimodal image editing, integrating visual content as part of the instruction to facilitate highly expressive and creative visual manipulations.

Recent text-to-image models produce high-quality images, yet text ambiguity hinders precise control when specific styles or objects are required. There have been a number of recent works dealing with learning and composing multiple objects and patterns. However, current work focuses almost entirely on image content, overlooking imaging factors such as camera lens, sensor types, imaging viewpoints, and scenes' domain characteristics. We introduce this new challenge as Imaging Factor Disentanglement and show limitations of current approaches in the regime. We, therefore, propose the new method Multi-factor disentanglement through Textual Inversion (MULTI). It consists of two stages: in the first stage, we learn general factors, and in the second stage, we extract dataset-specific ones. This setup enables the extension of existing datasets and novel factor combinations, thereby reducing distribution gaps. It further supports modifications of specific factors and image-to-image generation via ControlNets. The evaluation on our new DF-RICO benchmark demonstrates the effectiveness of MULTI and highlights the importance of Factor Disentanglement as a new direction of research.

Compositional generalization requires models to produce novel configurations from familiar parts. In diffusion models, prior compositional generation methods typically assume that the relevant concepts or conditioning signals are already available. We instead ask whether a pretrained diffusion model can discover query-specific concepts from the time-indexed scores it learns for the noisy marginals $p_t(x_t)$ and compose them at test time. Given a single out-of-distribution query, our method performs gradient ascent on $s_θ(x_t,t) \approx \nabla_{x_t}\log p_t(x_t)$ at multiple noising timesteps to recover local density modes, maps these modes into clean-space Gaussians, greedily selects relevant prototypes with a submodular likelihood objective, and combines them into a product-of-experts (PoE) teacher model with an analytic score. This teacher model can be sampled directly through classifier-free guidance or used to generate a sample pool for training a new class embedding and low-rank adapter. On held-out composition benchmarks built from ColorMNIST and CelebA, both the analytic PoE sampler and the low-rank adapted model outperform query-only and nearest trained-class baselines. These results suggest that the time-indexed score geometry of the diffusion model contains reusable density-mode concepts that support test-time compositional generation without a predefined concept library.

Diffusion models generate realistic visual content, yet often fail to produce rare but plausible compositions. When prompted with combinations that are valid but underrepresented in training data, such as a snowy beach or a rainbow at night, the generation process frequently collapses toward more common alternatives. We identify this failure mode as default completion bias, where denoising trajectories are implicitly attracted toward high-frequency semantic configurations. Existing guidance mechanisms do not explicitly model this competing tendency and therefore struggle to prevent such collapse. We introduce Default Completion Repulsion (DCR), a training-free framework that explicitly models and suppresses default completion behavior. DCR constructs a counterfactual attractor by relaxing the rare compositional factor while preserving surrounding semantics, inducing an alternative denoising trajectory reflecting the model's preferred completion. We define the discrepancy between target and attractor trajectories as a counterfactual drift, and propose a projection-based repulsion mechanism that removes guidance components aligned with this drift direction. This suppresses undesired frequent completions while preserving other semantic components. DCR operates entirely within the standard diffusion sampling process without retraining or architectural modification. Experiments on rare compositional prompts show that DCR improves compositional fidelity while maintaining visual quality. Our analysis further shows that the framework exposes and counteracts intrinsic model biases, offering a new perspective on controllable generation beyond explicit constraint enforcement.

Composition is a cornerstone of visual aesthetics, influencing the appeal of an image. While its principles operate independently of specific content, in practice, composition is often coupled with semantics. As a result, existing methods often enhance composition either through implicit learning or by semantics-based layout control, rather than explicitly modeling composition itself. To address this gap, we introduce Composer, a framework rooted in aesthetic theory, designed to model composition in a semantic-agnostic manner. First, it supports composition transfer by extracting key composition-aware representations from a reference image and leveraging a tailored conditional guidance module to control composition based on pre-trained diffusion models. Second, when users specify only text themes without a composition reference, Composer supports theme-driven composition retrieval by leveraging the in-context learning capabilities of Large Vision-Language Models (LVLMs), achieving explicit composition planning. To enhance composition in a reference-free mode, we conduct text-to-composition fine-tuning on the trained control module to enable implicit composition planning. Furthermore, we curated a high-quality dataset comprising 2 million image-text pairs using state-of-the-art generative models to support model training. Experimental results demonstrate that Composer significantly enhances aesthetic quality in text-to-image tasks and facilitates personalized composition control and transfer, offering users precision and flexibility in the creative process.

Text-to-image diffusion models achieve impressive visual fidelity, yet they remain unreliable in multi-object generation. Despite extensive empirical evidence of these failures, the underlying causes remain unclear. We begin by asking how much of this limitation arises from the data itself. To disentangle data effects, we consider two regimes across different dataset sizes: (1) concept generalization, where each individual concept is observed during training under potentially imbalanced data distributions, and (2) compositional generalization, where specific combinations of concepts are systematically held out. To study these regimes, we introduce mosaic (Multi-Object Spatial relations, AttrIbution, Counting), a controlled framework for dataset generation. By training diffusion models on mosaic, we find that scene complexity plays a dominant role rather than concept imbalance, and that counting is uniquely difficult to learn in low-data regimes. Moreover, compositional generalization collapses as more concept combinations are held out during training. These findings highlight fundamental limitations of diffusion models and motivate stronger inductive biases and data design for robust multi-object compositional generation.

Satellite image composition plays a critical role in remote sensing applications such as data augmentation, disaste simulation, and urban planning. We propose HarmoniDiff-RS, a training-free diffusion-based framework for harmonizing composite satellite images under diverse domain conditions. Our method aligns the source and target domains through a Latent Mean Shift operation that transfers radiometric characteristics between them. To balance harmonization and content preservation, we introduce a Timestep-wise Latent Fusion strategy by leveraging early inverted latents for high harmonization and late latents for semantic consistency to generate a set of composite candidates. A lightweight harmony classifier is trained to further automatically select the most coherent result among them. We also construct RSIC-H, a benchmark dataset for satellite image harmonization derived from fMoW, providing 500 paired composition samples. Experiments demonstrate that our method effectively performs satellite image composition, showing strong potential for scalable remote-sensing synthesis and simulation tasks. Code is available at: https://github.com/XiaoqiZhuang/HarmoniDiff-RS.

Narrative-driven product photography has become a prevalent paradigm in modern marketing, as coherent visual storytelling helps convey product value and establishes emotional engagement with consumers. However, existing image generation methods do not support structured narrative planning or cross-panel coordination, often resulting in weak storytelling and visual incoherence. In practice, narrative product photography is commonly presented as multi-grid collages, where multiple views or scenes jointly communicate a product narrative. To ensure visual consistency across grids and aesthetic harmony of the overall composition, we generate the collage as a single unified image rather than composing independently synthesized panels. We propose a self-reasoning agentic framework for narrative product grid collage generation. Given a product packshot and its name, the system first constructs a Product Narrative Framework that explicitly represents the product's identity, usage context, and situational environment, and translates it into complementary grids governed by a shared visual style. Constraint-aware prompts are then compiled and fed to a generation model that synthesizes the collage jointly. The generated output is evaluated on both content validity and photography quality, with explicit gates determining whether to proceed or refine. When evaluation fails, the system performs failure attribution and applies targeted refinement, enabling progressive improvement through iterative self-reflection. Experiments demonstrate that our framework consistently improves aesthetic quality, narrative richness, and visual coherence, compared to direct prompting baselines.

Graphic design creation involves harmoniously assembling multimodal components such as images, text, logos, and other visual assets collected from diverse sources, into a visually-appealing and cohesive design. Recent methods have largely focused on layout prediction or complementary element generation, while retaining input elements exactly, implicitly assuming that provided components are already stylistically harmonious. In practice, inputs often come from disparate sources and exhibit visual mismatch, making this assumption limiting. We argue that identity-preserving stylization and compositing of input elements is a critical missing ingredient for truly harmonized components-to-design pipelines. To this end, we propose GIST, a training-free, identity-preserving image compositor that sits between layout prediction and typography generation, and can be plugged into any existing components-to-design or design-refining pipeline without modification. We demonstrate this by integrating GIST with two substantially different existing methods, LaDeCo and Design-o-meter. GIST shows significant improvements in visual harmony and aesthetic quality across both pipelines, as validated by LLaVA-OV and GPT-4V on aspect-wise ratings and pairwise preference over naive pasting. Project Page: abhinav-mahajan10.github.io/GIST/.

Medical image generators trained on imbalanced data can fail at demographic intersections absent from training. We introduce CompDiff, which encodes age, sex and race separately and composes supervised demographic tokens alongside clinical text. Across chest radiographs and fundus images, CompDiff improves overall and subgroup fidelity relative to prompt conditioning (RoentGen-v2) and loss reweighting (FairDiffusion). It generalises in zero-shot generation to 16 chest X-ray intersections excluded from training, achieving the lowest mean FID-RadImageNet in every intersection. In a blinded reader study of these unseen intersections, two radiologists gave CompDiff the highest mean scores among generators for anatomical realism and agreement with the clinical impression, and selected its images most often as the most realistic. Pretraining with CompDiff images improved downstream classification, while CompDiff audit cohorts reduced estimation error on rare intersections. These findings support compositional demographic conditioning for extending medical image synthesis to underserved populations. Code: https://github.com/mahmoudibrahim98/CompDiff

While representation alignment with self-supervised models has been shown to improve diffusion model training, its potential for enhancing inference-time conditioning remains largely unexplored. We introduce Representation-Aligned Guidance (REPA-G), a framework that leverages these aligned representations, with rich semantic properties, to enable test-time conditioning from features, in generation. By optimizing a similarity objective (the potential) at inference, we steer the denoising process toward a conditioned representation extracted from a pre-trained feature extractor. Our method provides versatile control at multiple levels of granularity, ranging from patch level matching via single patches to broad semantic guidance using global image feature tokens. We further extend this to multi-concept composition, allowing for the faithful combination of distinct concepts. REPA-G operates entirely at inference time with no additional training required, offering a flexible and precise alternative to often ambiguous text prompts or coarse class labels. Our approach achieves high-quality, diverse generations on ImageNet and COCO. Code is available at https://github.com/valeoai/REPA-G

Existing personalized image generators still struggle to preserve multiple reference identities in natural and coherent multi-human generations. To address these limitations, we present CanvasComposer, an interactive framework for personalized group photo generation. Inspired by professional image-editing software, CanvasComposer allows users to place reference subjects on a shared canvas, where each subject keeps its own RGBA cutout of the input. This multi-reference canvas preserves reference content under overlap while providing an intuitive interface for organizing multiple identities; the subjects remain separate elements on the input canvas, and the model outputs a single personalized and harmonized image. To keep this representation efficient, transparent latent pruning retains only tokens from each subject's non-transparent region, and cross-reference training mitigates copy-paste artifacts by learning to harmonize references sampled from different images. Extensive experiments demonstrate that CanvasComposer achieves coherent generation and strong identity preservation compared to state-of-the-art methods in multi-human personalized image generation. Project page: https://snap-research.github.io/canvascomposer

Compositional generalization, the ability to generate novel combinations of known concepts, is a key ingredient for visual generative models. Yet, not all mechanisms that enable or inhibit it are fully understood. In this work, we conduct a systematic study of which design choices critically determine compositional generalization in image and video generation. By isolating independent design axes, we identify two key factors strongly associated with compositional success: (i) whether the training objective operates on a discrete or continuous distribution, and (ii) the completeness of conditioning information about constituent factors during training. We also show that relaxing the discrete loss with an auxiliary continuous latent objective can partially recover compositional performance in discrete models like MaskGIT. Our findings, corroborated by diverse compositional tasks and preliminary evidence in world models and LLMs, motivate a shift toward continuous objectives for compositional generalization.

Text-to-Image (T2I) diffusion models have made rapid progress on semantic alignment (generating what is described in the prompt), yet users still lack reliable control over aesthetic composition (how visual elements are put together). Prior work often treats aesthetics as a single, preference-driven notion (e.g., "high quality", "detailed", "breathtaking"), which does not map cleanly to compositional intent. We propose Aesthetic Alignment: aligning generated images to explicit, user-specified compositional constraints. We operationalize these constraints using the Principles of Art (PoA)-e.g., Balance, Rhythm, and Emphasis-commonly used in art education to describe composition. To support this task, we introduce CompArt, a dataset of 80,032 WikiArt images augmented with captions and PoA analyses produced by a multimodal LLM under structured prompting. We further propose ArtDapter, a lightweight and disentangled adapter that enables steering a pretrained T2I model along 10 PoA dimensions while retaining the base model's semantic capability. Experiments on CompArt show improved adherence to PoA controls over strong baselines under a dual evaluation protocol.