GeoDiT-Ω

GeoDiT-Ω enables precise synthesis using polygons, polylines, bounding boxes, points, along with text and geolocation conditioning. This unified framework supports hierarchical composition and adapts to both high and low annotation budgets.

Abstract

TL;DR — GeoDiT-Ω is a single diffusion model that generates satellite imagery directly from any geospatial primitive (polygons, polylines, boxes, or points), replacing task-specific pipelines with one model that boosts downstream segmentation, detection, road extraction, and classification.

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 GeoDiT-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, the flexibility of using primitives enables controllable synthetic data augmentation for multiple remote sensing tasks where labeled scenes are costly to acquire. Using a single generative model rather than task-specific pipelines, we improve performance on land-cover segmentation, object detection, road graph extraction, and scene classification.

Results

Qualitative Samples

Annotation complexity increases top to bottom; contextual modalities (text, geolocation) are added left to right. Polygons = building, polylines = road, boxes = recreational, points = trees.

Downstream Task Synthesis

GeoDiT-Omega synthetic samples on DIOR, OpenEarthMap, City-Scale, and AID

GeoDiT-Ω synthetic samples for four downstream tasks: (a) DIOR for object detection, (b) OpenEarthMap for land-cover segmentation, (c) City-Scale for road graph extraction, and (d) AID for scene classification.

Quantitative Comparison

Zero-shot evaluation on Git-Rand-15k and Git-Dense-3.5k. Bold is best, underline is second-best. T = text, L = geolocation, P = points, B = bounding box, Ω = {polygon, polyline, bounding box, point}.

Model	Condition	Git-Rand-15k			Git-Dense-3.5k
Model	Condition	FID↓	sFID↓	LPIPS↓	FID↓	sFID↓	LPIPS↓
General Models
InstanceDiffusion	T+B	94.18	22.60	0.5975	112.35	29.93	0.5979
GLIGEN	T+B	54.00	10.82	0.4792	96.28	24.94	0.5256
Geospatial Models
GeoSynth	T	45.59	18.88	0.5413	83.90	22.26	0.5149
VectorSynth	T+O	95.15	35.49	0.4441	72.26	25.87	0.4505
Text2Earth^*	T	25.93	5.09	0.4269	40.23	17.37	0.4871
GeoDiT-α-XL	T	14.21	5.13	0.3972	29.76	17.94	0.4853
GeoDiT-Σ-XL	T+P+L	12.01	5.09	0.3779	27.49	17.95	0.4692
GeoDiT-Ω-XL (Ours)	T+P+L	9.91	4.63	0.3612	20.57	17.44	0.3876
	T+B+L	9.72	4.45	0.3509	19.46	16.28	0.3594
	T+Ω+L	9.25	4.38	0.3438	18.20	16.21	0.3455

^* Text2Earth was trained on the entire Git-10M dataset (including our held-out test sets), which may inflate reported performance.

Method

Overview of GeoDiT-Ω architecture.

Geometry Aware Local Attention.

1

Encoding

Each scene instance can be described by any primitive format—polygon, polyline, box, or point. A Unified Primitive Encoder turns each into a shared token: geometry is Fourier-encoded and passed through an MLP, with the instance's text caption concatenated in. Global text and geolocation context are added separately—captions via LongCLIP, geolocation via RANGE.

2

Conditioning (GALA)

Unified primitive tokens condition generation through Geometry-Aware Local Attention (GALA): each token cross-attends to the diffusion transformer's visual tokens, then is modulated by a learned RBF field (MetaRBF+) and a spatial geometry field—together injecting a spatial inductive bias that adapts to each primitive's geometry. Global text conditions the model via standard cross-attention; geolocation and timestep via adaptive layer norm.

3

Generation

The conditioned tokens drive a standard diffusion transformer denoising process to produce the final satellite image.

BibTeX

@inproceedings{wei2026geoditomega,
  title     = {GeoDiT-Ω: Unified Spatial Control for Satellite Image
               Synthesis with Any Geospatial Primitive},
  author    = {Wei, Brian and Sastry, Srikumar and Cher, Daniel and Xing, Eric and Jacobs, Nathan},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}