PaperlayerView exact revision on arXiv

Source-linked AI summary

Genesis: A Generative Engine for Hierarchical Satellite Image Synthesis

Subash Khanal, Yangzhi Cui, Daniel Cher, Eric Xing, Brian Wei, Srikumar Sastry, Nathan Jacobs

arXiv:2609.02683v1cs.CV

TL;DR

Satellite-image generators do not jointly maintain consistency across zoom levels and neighboring tiles, despite the multi-scale structure of Earth observation. Genesis addresses this gap by combining quadtree super-resolution and mask-based outpainting, and evaluates the resulting task with dense500 and pyramid-level metrics. Its operators achieve state-of-the-art results on their respective subtasks, while full-pyramid evaluation compares model capacities and exposes the lack of an established end-to-end baseline.

  • Problem

    Existing satellite-image generators operate along only one axis, leaving complete pyramids jointly consistent across scale and space as an open challenge.

  • Method

    Genesis completes sparse quadtree seeds by composing a vertical super-resolution operator, a horizontal mask-based outpainting operator, and deterministic downsampling.

  • Results

    Genesis achieves state-of-the-art performance on both operator subtasks, with Genesis-H obtaining mean FID 12.6 on SR and 9.1 on OP.

  • Takeaways & Limitations

    The task, Genesis engine, dense500 benchmark, and pyramid-level metrics establish a framework for generating and evaluating seamless multi-resolution satellite maps from sparse seeds.

  • Takeaways & Limitations

    Because multi-scale tile completion is new and lacks an existing end-to-end method, comparisons with prior methods are limited to per-operator subtasks.

Abstract

from arXiv · show

Earth observation is fundamentally multi-scale; geospatial tasks span varied resolutions, and satellite imagery is organized into cascading tile pyramids that nest fine detail within wide coverage. Current generative models of satellite imagery, however, operate along a single axis: they either zoom to enhance a single tile's resolution or pan to extend imagery at a fixed scale. As a result, no existing method produces a complete pyramid that stays consistent across both scale and space, where a high-zoom tile must agree with the coarse context it refines and with the neighbors it meets. Motivated by this gap, we introduce a new task, multi-scale tile completion: given a sparse set of seed tiles at arbitrary zoom levels and positions, synthesize a complete, uniform quadtree that is globally consistent across both scale and space. We approach this task with Genesis, a generative engine that brings both axes together by composing two specialized operators over the quadtree: a vertical super-resolution model and a horizontal mask-based outpainting model, producing pyramids that are consistent across zoom levels and seamless across neighboring tiles. Each operator achieves state-of-the-art results on its subtask, and the engine propagates sparse seeds into seamless, multi-resolution maps from any initial configuration. To evaluate the task and benchmark Genesis, we introduce dense500, a fully observed multi-scale pyramid dataset spanning diverse geographic regions, together with a suite of pyramid-level metrics. Code, models, and our dataset are available at https://github.com/mvrl/genesis.

1 Introduction

Satellite imagery generation must address both the multi-resolution structure of tile pyramids and continuity across neighboring tiles. Genesis formalizes this as multi-scale tile completion and combines vertical super-resolution with horizontal outpainting, supported by the dense500 benchmark and pyramid-level metrics.

  • Motivation: Existing satellite-image generators operate either horizontally at fixed zoom or vertically across zoom levels, leaving joint scale-and-space consistency unresolved.The missing capability is agreement between high-zoom tiles and their coarse context while preserving continuity between neighbors.
  • Task: Multi-scale tile completion reconstructs every tile in a pyramid from sparse seeds at arbitrary positions and zoom levels under vertical and horizontal consistency constraints.The task targets complete, uniform quadtrees rather than isolated generated images.
  • Method: Genesis composes vertical super-resolution and horizontal mask-based outpainting over the quadtree to expand arbitrary seed configurations into seamless, multi-resolution maps.The framework unifies the two spatial directions through complementary generative operators.
  • Evaluation: The paper introduces dense500, a fully observed multi-scale pyramid benchmark, with pyramid-level metrics for evaluating multi-scale generation quality.The benchmark and protocol are intended to support future research on multi-scale generative modeling.

2 Related Work

Satellite-image generation builds on general image-generation advances while requiring domain-specific conditioning and spatial operators. Super-resolution addresses refinement across scale, whereas inpainting and outpainting complete content within a fixed zoom, typically one image at a time.

  • General image generation: Recent image-generation systems have advanced through diffusion, latent representations, transformer backbones, and flow-based training.These developments are presented as the broader technical context for modern image synthesis.
  • Satellite image generation: Satellite-image generation adapts conditioning principles to overhead imagery’s distinct perspective, scale, spectral content, and metadata.The related work spans text-to-image generation and other domain-specific conditioning approaches.
  • Super-resolution: Single-image super-resolution has progressed from CNN mappings through perceptual, adversarial, attention-based, and diffusion-based methods.SwinIR is identified as a strong perceptual baseline, while FastDiffSR accelerates diffusion-based remote-sensing super-resolution.
  • Inpainting and outpainting: Inpainting and outpainting methods hallucinate plausible masked or out-of-frame content from visible context, but conventional methods complete a single image.Covering regions larger than one tile requires assembling multiple tiles rather than completing only one image.

3 Data

Genesis is trained on sparse global satellite imagery and evaluated with dense500, which provides fully observed, geographically diverse quadtrees. The data construction enforces disjoint evaluation and filters sites to preserve valid multi-level imagery across several resolution ranges.

  • Datasets: The models are trained on Git-10M, while dense500 supplies 500 fully observed depth-4 quadtrees for evaluating multi-scale completion.Both datasets use 256×256 tiles arranged in quadtree structures with doubling ground resolution at finer zooms.
  • Training data: Git-10M contains approximately 10M satellite tiles across zoom levels 10–18, but most georeferenced tiles lack some relatives and require partial hierarchy reconstruction.Approximately 7M tiles carry Web-Mercator coordinates and participate in the quadtree hierarchy.
  • Splits: Evaluation quads are held out as complete subtrees, excluding every tile within each evaluation subtree from training to prevent test-content leakage.The spatial test split equally samples urban and non-urban quads.
  • Benchmark design: dense500 contains 500 depth-4 quadtrees with 85 tiles each, organized into five zoom windows spanning roughly 76 to 0.6 m/px.Because every tile is observed, arbitrary seed configurations can be completed and scored at every pyramid level.
  • Site selection: Sites are population-weighted, urban-restricted, continent-stratified, geographically spaced, disjoint from training, and filtered when imagery sources shift across adjacent zoom levels.The final filtering addresses satellite-to-aerial transitions that occur at higher zoom levels.

4 Method

Genesis formulates multi-scale tile completion as filling a quadtree from arbitrary seed tiles while enforcing vertical parent–child agreement and horizontal continuity. It combines super-resolution, downsampling, and mask-based outpainting, scheduled to propagate seeds into a complete pyramid.

  • Multi-scale Tile Completion: Multi-scale tile completion fills every missing tile in a target quadtree from seeds at arbitrary zoom levels and positions.The result must satisfy vertical consistency between parents and children and horizontal continuity between adjacent tiles.
  • Quadtree Operators: Genesis uses super-resolution to map an N×N tile at zoom z to the 2N×2N mosaic of its four children at zoom z+1.Downsampling provides the deterministic upward operator, reconstructing a parent from its children.
  • Quadtree Operators: Mask-based outpainting completes a tile from an arbitrary subset of known quadrants, providing the horizontal completion operator.During inference, known pixels remain fixed while only the masked hole is generated.
  • Pixel-space Generative Transformer: Both operators are independently trained JiT flow-matching transformers operating directly in pixel space.Flow matching interpolates clean images with Gaussian noise, while Genesis predicts clean images and applies a velocity-space loss.
  • Pyramid Scheduling: Genesis concentrates generative work at one working level, then applies successive super-resolution passes upward and downsampling passes downward.The working level is chosen r levels above the leaves, balancing the outpainting grid size against high-resolution refinement.
  • Pyramid Scheduling: The engine brings seeds to the working level, completes missing regions by maximum-coverage outpainting, refines to leaves, repairs seams, and reconciles coarser tiles.Coarser seeds are super-resolved, finer seeds are downsampled into quadrants, and leaf seeds are anchored during refinement.

5 Experimental Details

The experiments specify curricula, sampling procedures, operator baselines, and dense500 pyramid evaluation under three seed protocols. Pyramid assessment combines image quality, semantic alignment, and cross-scale consistency metrics.

  • Backbone and Training: Both operators use JiT-H or JiT-B backbones and four-stage easy-to-hard curricula, with transitions at 20/40/60% of training.SR curriculum stages vary parent zoom, while OP stages vary the mask regime.
  • Operator Evaluation: Per-operator evaluation compares Genesis SR with FastDiffSR, SwinIR, and ZoomLDM, and Genesis OP with SD2-Inpaint, SD2-Inpaint+Text, and Text2Earth-Inpaint.SR uses Git-10M parent-to-child scoring; OP uses shared hole-aware metrics and boundary difference.
  • Pyramid Generation: The pyramid engine uses whole-mosaic MultiDiffusion-style super-resolution with overlapping windows and re-anchors leaf seeds at every step.The SR pass is warm-started from a bicubic upsample at τ=0.4 to avoid independently generated tile seams.
  • Pyramid Generation: Maximum-coverage outpainting fills the penultimate level using batched, cell-disjoint 256×256 windows over 128×128 quadrants.Completed tiles are written back as soon as their four quadrants are resolved.
  • Pyramid Generation: Seam repair ranks internal leaf boundaries by mean pixel discontinuity and repairs the worst boundaries with masked outpainting bands.The procedure uses threshold 14 and repairs up to 64 boundaries.
  • Pyramid Evaluation: dense500 evaluation reports image quality, semantic alignment, and pyramid-consistency metrics across three seed protocols applied to all 500 subtrees.The protocols use one, three, or four spatially independent seeds, testing increasingly dense and varied anchoring conditions.

6 Results

Genesis performs strongly on both super-resolution and outpainting, then composes these operators for consistent full-pyramid generation on dense500. Results also show that perceptual and boundary metrics are especially relevant when operations are chained across scales and neighboring tiles.

  • Super-resolution: Genesis-H achieves the best perceptual and distributional super-resolution scores across both test splits and all six split/zoom groups.Its FID ranges from 9.2–14.8, compared with 32–47 for SwinIR and 42–70 for ZoomLDM.
  • Super-resolution: Pixel-fidelity metrics favor coarse children, while perceptual differences from baselines are widest at fine zoom, showing that PSNR and SSIM alone miss perceptual quality.At z18, fine texture dominates; bicubic can score highly on PSNR and SSIM despite blurry outputs penalized by LPIPS and FID.
  • Outpainting: Genesis-H attains the lowest FID, LPIPS, and B-diff in all eight outpainting split/regime cells against three compared baselines.Outpainting becomes harder as the missing region grows: FID rises from about 3.2 at quad3 to 17 at quad1.
  • Summary across Regimes: Across aggregated regimes, Genesis-H leads both subtasks, with mean FID 12.6 for super-resolution and 9.1 for outpainting.The results identify one model family as strongest on both operators used by the engine.
  • Full Pyramid Generation: The full-pyramid comparison lacks an existing end-to-end method, so Genesis is compared with prior methods on its operator subtasks and with internal capacities on dense500.Qualitative completions show spatial and hierarchical consistency, while failures are concentrated in the single-seed extreme.
  • Full Pyramid Generation: On dense500, Genesis-H improves over Genesis-B on nearly every full-pyramid metric, including FID 72.0 vs. 120.1 and CLIP-I 0.843 vs. 0.800.The improvement trades additional compute for generative quality.
  • Full Pyramid Generation: Denser seeding improves FID and semantic alignment, while LR-PSNRbox remains high and roughly constant across protocols.For Genesis-H, FID falls from 99.4 with one seed to 62.2 with four seeds, while LR-PSNRbox indicates stable parent–child agreement.

7 Conclusion

The paper introduces multi-scale tile completion and Genesis, which synthesizes complete satellite-image pyramids consistent across geographic space and zoom levels. It also contributes dense500 and pyramid-level metrics for evaluating this task.

  • Multi-scale tile completion synthesizes complete satellite-image pyramids from sparse observations while maintaining consistency across geographic space and zoom levels.
  • Genesis combines a vertical super-resolution model with a horizontal mask-based outpainting model to complete full pyramids from any seed configuration.
  • dense500 is a fully observed multi-scale pyramid benchmark spanning diverse geographic regions, accompanied by pyramid-level metrics.
  • The evaluation shows that Genesis produces realistic pyramids that remain semantically aligned with seeds and consistent across scales.
  • The completed pyramids support applications including virtual environments, urban-layout prototyping, and multi-resolution remote-sensing training data.

A.1 Data Splits

The georeferenced Git-10M subset supplies evaluation splits built from complete quads, while the broader training set also includes partial quads. Geographic coverage is near-global, and Table 7 summarizes the splits.

  • Complete quads, whose four z+1 children are present, form the population for the evaluation splits.
  • The evaluation uses disjoint spatial and random test/validation sets, with spatial splits balanced 50/50 between urban and non-urban tiles.
  • The training set contains the remaining tiles and additionally includes partial quads with one to three children.
  • Git-10M training tiles provide near-global geographic coverage, while sites are shown separately in red.
  • Table 7 reports the data splits over the georeferenced Git-10M subset.

A.2 dense500 Site Selection and Filtering

dense500 sites are selected from populated locations and filtered to exclude overzoomed or visually flat imagery. A site is retained only when all 85 tiles pass these checks.

  • Site centers are sampled from the LandScan 2024 population raster, with urban sites restricted to cells above the 70th percentile of populated cells.
  • Tiles are rejected when they are overzoomed or flat, detected respectively through low Laplacian variance or low per-channel standard deviation.
  • A site is kept only if all 85 tiles pass filtering, so published sites contain no upsampled or empty imagery.

A.3 Metric Details

This section provides additional details about the metrics used to evaluate generated pyramids.

  • The paper provides additional details on the metrics used for evaluating generated pyramids.
  • The metric details concern evaluation of generated pyramids.
  • The section supplements the paper’s evaluation protocol with further metric information.

A.4 Semantic Alignment

The evaluation combines semantic alignment metrics with pyramid-consistency measures to assess whether generated satellite regions match real scenes and remain coherent across scales.

  • Semantic Alignment: Semantic alignment is evaluated at category, instance, and scene levels using CLIP-I, DINOv3-sat, and caption agreement.CLIP-I and DINOv3-sat compare embedding cosine similarity, while caption agreement compares sentence embeddings of generated and real-region captions.
  • Semantic Alignment: Caption agreement captions synthesized and real regions with Qwen2.5-VL-3B-Instruct, then compares their Sentence-BERT representations.The reported score is the cosine similarity between the two caption embeddings.
  • Pyramid Consistency: LR-PSNR measures intra-scale self-consistency between a generated parent and its four generated children without requiring ground truth.The child tiles are stitched into a mosaic and downsampled with the physically correct box operator; lower LR-PSNR indicates greater disagreement.
  • Pyramid Consistency: xLR-PSNR compares the generated child mosaic with the ground-truth child mosaic at the parent’s resolution, separating fidelity from self-consistency.This addresses cases where parent and children agree internally but are jointly incorrect.
  • Pyramid Consistency: RAPSD spectral distance compares generated and ground-truth child mosaics through mean absolute log-power differences in radially averaged power spectra.The radial frequency representation uses azimuthally averaged power S(f).
Loading 2609.02683v1…