TerraDiT-$Ω$: Unified Spatial Control for Satellite Image Synthesis with Any Geospatial Primitive

2026-06-29

AI Summary

Existing methods for satellite image generation struggle to effectively leverage diverse geospatial vector primitives—such as polygons, polylines, and points—to achieve high-fidelity, controllable synthesis. This work proposes the first unified spatial control framework based on a diffusion transformer architecture, incorporating a geometry-aware local attention mechanism that directly ingests arbitrary native vector primitives as conditioning inputs. By explicitly injecting geometric information, the model enables multi-granular spatial layout control. The approach consistently outperforms existing dense or sparse control methods across various vector conditions, achieving high-quality, controllable image generation with a single unified model. Furthermore, it significantly enhances performance on downstream tasks, including land cover segmentation, object detection, road extraction, and scene classification.

Abstract

Generative models have achieved remarkable progress, yet applying them to satellite imagery remains challenging. Unlike natural imagery, satellite scenes are structured by spatially complex and semantically distinct geometries. Prior work addresses this complexity by adapting natural image frameworks using dense rasters or sparse prompts, trading off annotation cost and fidelity while breaking compatibility with vector primitives commonly used to represent geographic information. We introduce TerraDiT-$Ω$, a unified spatial control framework that generates satellite imagery directly from any native geospatial primitive. By jointly leveraging precise annotations (polygons, polylines) and coarser ones (bounding boxes, points), the model supports controllable layouts across varying annotation budgets, broadening applicability to design tasks such as urban planning while remaining naturally compatible with end-to-end GeoAI workflows. To effectively leverage these primitives during generation, we propose Geometry-Aware Local Attention, a conditioning mechanism that injects explicit geometric cues into the attention space. Across all conditioning formats, our approach consistently outperforms both dense-control and sparse-control baselines. Furthermore, this flexibility enables controllable synthetic data augmentation using a single generative model, improving downstream performance on land-cover segmentation, object detection, road graph extraction, and scene classification. Code, data, and weights are available at https://github.com/mvrl/TerraDiT.

Problem

• satellite image synthesis
• geospatial primitives
• spatial control
• vector representation
• generative modeling

Innovation

• geospatial primitives
• satellite image synthesis
• Geometry-Aware Local Attention
• unified spatial control
• controllable generation