Layout generation, which predicts the spatial configuration of objects such as bounding boxes and categories, serves as a fundamental structural blueprint for controllable image synthesis. In real-world scenarios, object placement is not arbitrary. Ra...
Layout generation, which predicts the spatial configuration of objects such as bounding boxes and categories, serves as a fundamental structural blueprint for controllable image synthesis. In real-world scenarios, object placement is not arbitrary. Rather, humans determine optimal locations by considering semantic and spatial relationships with surrounding anchor objects (e.g., placing a mouse to the right of the keyboard). However, existing inpainting models primarily rely on pixel-level statistical patterns, often failing to synthesize realistic content when the target object is generated in a physically implausible location. Moreover, most layout generation models focus on synthesizing entire scenes from scratch, lacking the flexibility to insert target objects while preserving the existing visual context. Furthermore, they are often confined to pre-defined closed-set classes, limiting their capability to handle diverse natural language queries. To address these limitations, we propose ArcLayout, a context-aware framework employing a large-scale open-vocabulary approach via Large Language Models (LLMs). By leveraging the rich semantic embedding space of LLMs, our model handles unrestricted object categories and explicitly models spatial dependencies relative to anchor objects to infer the optimal layout. Through experiments, we demonstrate that our model effectively captures intrinsic spatial commonsense. Specifically, utilizing the generated layouts as structural guides significantly improves image realism and object recognizability, achieving Top-1 accuracies of 64.13% for position (Top-5 83.1%) and 63.66% for shape (Top-5 79.4%), and recording an FID score of 62.02 with the Paint-by-Example inpainting model. These results underscore that explicit structural design is a prerequisite for high-fidelity image generation.