Research map / Image generation / Autoregressive and latent generation
TerraDiT-Ω: Unified Spatial Control for Satellite Image Synthesis with Any Geospatial Primitive
Brian Wei, Srikumar Sastry, Daniel Cher, Eric Xing, Nathan Jacobs
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add rotation θ to the RBF prior (ablation) | GALA (B, 400k) · Ablation split | FID: 23.52 -> 22.41 | Improved |
| add spatial geometry field modulation (ablation) | GALA (B, 400k) · Ablation split | FID: 22.41 -> 21.97 | Improved |
| add RBF Gaussian prior to cross attention (ablation) | Unified Primitive Encoder (B, 400k) · Ablation split | FID: 24.08 -> 23.52 | Improved |
| add primitives Ω as conditioning (ablation) | TerraDiT-Ω (B, 400k) · Ablation split | FID: 27.19 -> 21.97 | Improved |
| add global text conditioning (ablation) | TerraDiT-Ω (B, 400k) · Ablation split | FID: 24.90 -> 21.97 | Improved |
| add geolocation conditioning (ablation) | TerraDiT-Ω (B, 400k) · Ablation split | FID: 23.53 -> 21.97 | Improved |
| use GALA instead of Gated Self-Attention (ablation) | Unified Primitive Encoder (B, 400k) · Ablation split | FID: 22.27 -> 21.97 | Improved |
| use GALA instead of Instance Masked Attention (ablation) | Unified Primitive Encoder (B, 400k) · Ablation split | FID: 22.31 -> 21.97 | Improved |
| richer primitives: points to box to combined Ω (ablation) | TerraDiT-Ω-XL · Git-Rand-15k | FID: 9.91 -> 9.72 -> 9.25 | Improved |
| richer primitives from points to Ω (ablation) | TerraDiT-Ω-XL · Git-Rand-15k CAS | Top-1 accuracy: 76.66 -> 79.36 | Improved |
| add x2 synthetic data | TerraDiT-Ω-XL on DIOR · DIOR + RT-DETR | mAP@50-95: 56.14 -> 55.90 | Mixed |
| add AeroGen synthetic data without filtering | RT-DETR on DIOR · DIOR + RT-DETR | mAP@50-95: 55.15 -> 47.20 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- InstanceDiffusion — InstanceDiffusion: Instance-level Control for Image Generation (CVPR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Geospecific View Generation - Geometry-Context Aware High-resolution Ground View Inference from Satellite Views (ECCV 2024 · similarity 0.89)
- Controllable Satellite-to-Street-View Synthesis with Precise Pose Alignment and Zero-Shot Environmental Control (ICLR 2025 · similarity 0.88)
- DiffusionSat: A Generative Foundation Model for Satellite Imagery (ICLR 2024 · similarity 0.87)
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.86)
- EcoMapper: Generative Modeling for Climate-Aware Satellite Imagery (ICML 2025 · similarity 0.86)
- Satellite to GroundScape - Large-scale Consistent Ground View Generation from Satellite Views (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
