Genesis: Hierarchical Satellite Image Synthesis

The task: multi-scale tile completion

TL;DR — Given sparse seed tiles at arbitrary positions and zoom levels, synthesize the full quadtree such that the output preserves every seed, is seamless across neighboring tiles, and stays consistent across zoom levels.

Sparse seed tiles are completed into a full multi-scale quadtree — Seed tiles (red) can sit anywhere in the pyramid; Genesis fills everything else.

How Genesis works

Two flow-matching JiT operators — a parent-conditional super-resolution model (one 256×256 parent → the 512×512 mosaic of its four children) and a mask-based outpainting model — composed into one pyramid engine.

The Genesis pyramid-completion pipeline — (1) take the seeds → (2) bring them to a common working level → (3) greedily outpaint to complete that level → (4–5) super-resolve down to the leaf tiles → (6) fill the coarser levels by downsampling — consistent across scales, honoring every seed.

Results

Cross-scale synthesis from seed tiles — Zooming from z=15 to z=18 through a generated pyramid.

Pyramid completions across many sites — Completions across distinct sites (columns) and zoom levels (rows): outputs are consistent vertically across zoom levels and horizontally within each level.

Code, models & reproducibility

🌍 Code — github.com/mvrl/genesis: training, evaluation, the pyramid engine, and interactive Gradio demos.

🤗 Models — huggingface.co/MVRL/genesis: all checkpoints (SR + outpainting, JiT-B/16 & JiT-H/16), the DINOv3-sat encoder, and the evaluation metadata.

🗺️ Benchmark — dense500: 500 fully-observed depth-4 quadtrees (85 tiles each) across five zoom windows, for scoring pyramid completion against ground truth at every level.

Citation

@inproceedings{khanal2026genesis,
  title     = {{Genesis}: A Generative Engine for Hierarchical Satellite Image Synthesis},
  author    = {Khanal, Subash and Cui, Yangzhi and Cher, Daniel and Xing, Eric and
               Wei, Brian and Sastry, Srikumar and Jacobs, Nathan},
  booktitle = {ACM SIGSPATIAL International Conference on Advances in
               Geographic Information Systems (SIGSPATIAL)},
  year      = {2026},
}