Original title: Genesis: A Generative Engine for Hierarchical Satellite Image Synthesis
Subash Khanal, Yangzhi Cui, Daniel Cher, Eric Xing +3 · September 2026
Satellite map services organize Earth pictures into tile pyramids. Wide views nest blurry landscapes, while zoomed-in views show sharp details. Today, generative image tools only work along one direction. They can zoom in on one picture to add sharp details, or they can pan across to stretch one picture sideways. None can build a full pyramid where views match their neighbors and look right at every zoom level. To fix this, researchers created Genesis. Genesis starts with just a few scattered sample tiles at any location or zoom level. It then uses two tools over a branching pyramid structure called a quadtree. One tool zooms down vertically to increase image sharpness. The other tool expands outward horizontally to fill in missing side-by-side views. Together, these tools build complete, seamless, multi-resolution maps that stay fully consistent across both zoom levels and surrounding spaces from any starting setup.
You have a puzzle with only a few corner pieces placed. The middle is missing, so you can't see how the image continues. Here, the system starts with a few tiles at different zoom levels and fills in all the missing parts so that the whole image is complete and consistent.
This matters for earth observation tasks like mapping and monitoring, where consistent views across different zoom levels are needed. It allows generating complete imagery from sparse data, making large-area analysis more practical.
The abstract does not specify any limitations of the work, but it notes that previous methods operated along single axes.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Genesis: A Generative Engine for Hierarchical Satellite Image Synthesis — Subash Khanal, Yangzhi Cui, Daniel Cher et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.