9.24B · Flow Matching DiT · Earth Observation

A geo-aware generative foundation model for Earth observation.

Trained from scratch on EO data with native text, ground sample distance, latitude, and longitude conditioning.

Jeonghyeok Do·Munchurl Kim·KAIST

Read the paper Code ↗Model weights ↗

Selected zero-shot observations

10 / 10

Houses along both sides of a highway — 06

White rocks and brown mountain earth — 08

01Overview

A foundation prior native to the overhead view.

Existing EO generators often inherit the perspective and composition biases of natural-image diffusion models. GeoCore-9B instead trains a large Flow Matching-based Diffusion Transformer from scratch on satellite RGB imagery.

Local text tokens join continuous GSD, latitude, and longitude metadata. A training-only Geospatial Semantic Alignment objective stabilizes structure, and the resulting prior transfers to cloud removal and SAR-to-optical translation through parameter-efficient adaptation.

01 / Native scale

Built from EO data

The DiT backbone is initialized and trained from scratch on Git-10M.

02 / Geographic context

Conditioned by place

Text, GSD, latitude, and longitude jointly guide generation.

03 / Practical transfer

Beyond proxy synthesis

LoRA adaptation supports restoration and cross-modal translation.

Architecture8 double + 24 single blocks

Hidden / heads4096 / 32

Training300K iters · batch 1024

Output256 × 256 RGB optical

02Method

Flow matching, anchored by geospatial semantics.

A scalable latent DiT learns the EO distribution while a satellite-specialist teacher supplies structural guidance only during pre-training.

GeoCore-9B training and inference architecture — Fig. 01 · System map
Text features enter the DiT alongside GSD and geographic coordinates. The frozen DINOv3-Sat teacher and projection head are removed at inference.

Geo-conditioned DiT

T5 latent tokens and CLIP global text join sinusoidally encoded GSD, latitude, and longitude through AdaLN.

Flow Matching

The model predicts velocity along a linear path from Gaussian noise to the VAE latent.

Semantic alignment

Block 8 aligns with frozen DINOv3-Sat dense features at μ = 0.5, with zero inference overhead.

03Geo-aware generation

Describe the scene. Set the scale. Locate it on Earth.

Language controls semantics, GSD controls physical granularity, and coordinates retrieve broad geographic priors.

03.A / Text

Text remains orthographic.

GeoCore-9B follows prompts for infrastructure, terrain, and settlement without drifting into ordinary photographic composition.

Dense built environments
Natural land cover
Fine man-made geometry

Text-conditioned comparison with CRS-Diff and Text2Earth — Fig. 02 · Text control
Cleaner EO structure across diverse scene descriptions.

03.B / GSD

One scene, many physical scales.

Select a study set to inspect the transition from fine local detail to broad land cover.

Full scale GSD comparison — Fig. 03 · Full scale study
GeoCore-9B changes visual granularity across 1–32 m GSD while preserving scene identity.

03.C / Coordinates

Coordinates alone recover geographic priors.

Without text, the model produces distinct terrain priors for New York City, the Sahara, Antarctica, and the Amazon.

Coordinate-only generations for four world regions — Fig. 04 · Coordinate-only control
Text is removed; only latitude and longitude are supplied.

04GSA study

Sharper structure, faster convergence, no inference tax.

The matched 9B ablation changes only the Geospatial Semantic Alignment weight.

Generation comparison without and with GSA — Fig. 05 · Structural fidelity
GSA reduces fragmented boundaries in farms, roofs, roads, and roundabouts.

FID curves with and without GSA — Fig. 06 · Training trajectory
Lower FID throughout pre-training on a fixed 10K Git-10M subset.

RSICD FID: −9.61
SAR translation FID: −7.87
Cloud removal PSNR: +1.256 dB

The teacher and projection head supervise pre-training only. Both are discarded before downstream adaptation and inference.

05Results

A generative prior that transfers to practical EO tasks.

GeoCore-9B is adapted with LoRA while its pre-trained backbone remains frozen.

Task / dataset	Metric 01	Metric 02	Metric 03	Metric 04
Text-to-imageRSICD	22.15IS ↑	18.82FID ↓ · best	27.15CLIP ↑	—
Cloud removalSen2-MTC	20.809PSNR ↑ · best	0.799SSIM ↑ · best	0.256LPIPS ↓	—
SAR-to-opticalQXS-SAROPT	12.05FID ↓ · best	0.377LPIPS ↓	0.0163HF-SCC ↑ · best	0.370SSIM ↑

05.A / Practical adaptation

Clouds out. Structure retained.

GeoCore-9B reconstructs roads, fields, and boundaries beneath severe cloud contamination.

Cloud removal comparison set one — Study set 01
Input, specialist baselines, GeoCore-9B, and ground truth.

Cloud removal comparison set two — Study set 02
Additional qualitative comparisons under the same protocol.

06Resources

Read, reproduce, and build on GeoCore-9B.

Manuscript

Paper PDF

Method, controlled analyses, results, and limitations.

↗ Implementation

GitHub repository

Inference, models, evaluation, configurations, and scripts.

↗ Checkpoint

Hugging Face

Model card, configuration, and released weights.

↗

Quick start

Generate from text and geo metadata.

git clone https://github.com/KAIST-VICLab/GeoCore-9B
cd GeoCore-9B && pip install -r requirements.txt

python inference.py \
    --ckpt /path/to/GeoCore-9B --vae /path/to/ae.safetensors \
    --prompt "A satellite view of a highly dense urban city with towering skyscrapers" \
    --lon 126.97 --lat 37.56 --res 0.0 \
    --num-samples 4 --out samples/

Citation

Cite GeoCore-9B.

@article{do2026geocore,
  title   = {GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation},
  author  = {Do, Jeonghyeok and Kim, Munchurl},
  year    = {2026}
}

Acknowledgment

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) under the Sejong Science Fellowship Program (RS-2026-25484549), for the project “Visualizing the Invisible Earth: A Reliability-Aware All-in-One SAR Analysis Framework with Foundation Models.”