Research map / Image generation / Fast diffusion sampling

GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Rectified flow straightening · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains a 9B Flow Matching DiT from scratch on the Git-10M EO dataset with text and GSD/lat/lon conditioning, adds a training-only alignment loss to a frozen DINOv3-Sat teacher, and adapts it with LoRA to cloud removal and SAR-to-optical translation.

Problem

Generative models for Earth observation typically fine-tune natural-image diffusion priors, which introduce perspective biases and limit scale, and existing EO evaluations rely on proxy generation tasks rather than practical restoration or cross-modal translation.

Approach

A 9B DiT trained from scratch on EO data with continuous geospatial metadata conditioning (GSD, latitude, longitude), stabilized by a Geospatial Semantic Alignment loss that aligns intermediate DiT features to a frozen satellite-specialist teacher (DINOv3-Sat) during training only.

Key results

• On Sen2-MTC cloud removal, GeoCore-9B reaches 20.809 PSNR and 0.799 SSIM, above the best specialist EMRDM (20.067 PSNR, 0.709 SSIM).
• On QXS-SAROPT SAR-to-optical translation, GeoCore-9B achieves FID 12.05 versus 18.15 for the best specialist C-DiffSET.
• In a matched 9B ablation, GSA reduces RSICD FID from 28.43 to 18.82 and QXS-SAROPT FID from 19.92 to 12.05.

Surprising findings

• GeoCore-9B's LPIPS on cloud removal (0.256) is not best; EMRDM achieves 0.255 and the paper says LPIPS is comparable to the strongest specialist.
• Coordinate-only generations found no near-duplicates in the 10.5M-image Git-10M corpus under the stated global-feature threshold.

Limitations

• Relies on a natural-image VAE, which bounds reconstruction fidelity; lower HF-SCC on Sen2-MTC.
• Only RGB optical output at 256x256; does not generate SAR or multispectral imagery.
• Matched ablations do not disentangle model scale, data, compute, and GSA; no matched w/o-GSA probe for transfer experiments.

Experimental setup

Datasets: Git-10M RSICD Sen2-MTC QXS-SAROPT EuroSAT LoveDA BRIGHT

Benchmarks: RSICD Sen2-MTC QXS-SAROPT EuroSAT LoveDA BRIGHT

Metrics: FID IS CLIP score PSNR SSIM LPIPS HF-SCC top-1 accuracy mIoU

Baselines: CRS-Diff Text2Earth DiffusionSat McGAN DiffCR EMRDM UnCRtainTS C-DiffSET cBBDM ControlNet

Builds on: Flow Matching DiT REPA CLIP T5

Closest papers by meaning

• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.88)
• REOBench: Benchmarking Robustness of Earth Observation Foundation Models (NeurIPS 2025 · similarity 0.88)
• GeoDiT: A Diffusion-based Vision-Language Model for Geospatial Understanding (CVPR 2026 · similarity 0.88)
• GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding (CVPR 2026 · similarity 0.88)
• Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation (CVPR 2025 · similarity 0.87)
• No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data (ICLR 2025 · similarity 0.87)