SatSplatDiff: Geometry-preserving generative refinement for high-fidelity satellite Gaussian Splatting

Abstract: Gaussian Splatting has been recently explored for satellite 3D reconstruction, demonstrating flexibility and efficiency in representing radiometrically diverse satellite scenes. However, the limited top viewpoint of satellite imagery results in insufficient supervision on building facades, leaving surface holes and degraded visual fidelity. Generative refinement, which leverages pretrained generative priors to iteratively refine and update the rendered images used as supervision targets, has recently been investigated to improve the visual fidelity of Gaussian-rendered images. However, since these models refine each view independently, the resulting images can generate hallucinations and break photo-consistency, leading to geometric degradation. To address these limitations, we propose SatSplatDiff, which aims to minimize geometric degradation prevalent in generative refinement. Building on photogrammetric DSM initialization and 2DGS-based shadow casting established in our prior work SatSplat, we first introduce monocular depth supervision and multi-scale geometric refinement to establish a geometrically accurate and well-regularized surface representation. We then apply shadow-guided generative refinement, where geometrically calculated shadow maps guide the Gaussians to maintain consistency with the underlying geometry, improving visual fidelity while reducing geometric degradation. Extensive evaluations on the IARPA2016 and DFC2019 datasets demonstrate state-of-the-art performance, reducing geometric MAE by up to 18% and improving visual fidelity (FID-CLIP) by 28-45% over existing baselines. Our method delivers up to 5x resolution enhancement with minimal hallucination and sensor-consistent appearance, demonstrating seamless cross-tile consistency and strong scalability for large-scale reconstruction. Source code is available at https://github.com/GDAOSU/SatSplatDiff

• The paper presents a novel approach that integrates DSM-based bootstrapping, multi-scale geometric optimization, and shadow-guided diffusion to achieve high-fidelity 3D reconstruction.
• It demonstrates significant improvements in geometric accuracy and visual fidelity, with sub-meter error reduction and up to 5× resolution enhancement.
• The methodology preserves geometric consistency while leveraging diffusion models to refine appearance without hallucinating structures, enabling scalable urban mapping.

Geometry-Preserving Generative Refinement for Satellite 3D Gaussian Splatting

Introduction

"SatSplatDiff: Geometry-preserving generative refinement for high-fidelity satellite Gaussian Splatting" (2606.27223) addresses the persistent challenge of reconstructing high-fidelity 3D urban models from satellite images using Gaussian Splatting (GS). Traditional GS pipelines demonstrate flexibility and computational efficiency for large-scale scene representation. However, in the context of satellite imagery—characterized by limited nadir-centric viewpoints and radiometric complexity—such methods struggle to reconstruct occluded vertical surfaces like building facades, leading to incomplete geometry and degraded photo-realism. Generative refinement strategies leveraging high-capacity diffusion priors have recently been proposed for image-based novel view refinement, but they often undermine geometric integrity by producing hallucinated content lacking multi-view consistency, which is critical for photogrammetric fidelity.

SatSplatDiff systematically integrates multi-modal geometric constraints and a novel shadow-guided generative refinement technique to overcome these deficiencies. The proposed pipeline leverages photogrammetric DSM initialization, monocular depth supervision, multi-scale geometric optimization, adaptive Gaussian densification, and a shadow-conditioned generative refinement module driven by diffusion models. Comprehensive experiments on the IARPA2016 and DFC2019 datasets demonstrate superior geometric and photometric performance, along with strong cross-tile consistency and up to $5\times$ resolution enhancement.

Methodological Innovations

Photogrammetric Initialization and DSM-based GS Bootstrapping

Unlike prior NeRF/GS approaches relying on sparse point clouds or volumetric Gaussian initialization, SatSplatDiff builds on a dense photogrammetric DSM, obtained via state-of-the-art stereo pipelines and bundle adjustment. Gaussian primitives are initialized directly from the DSM grid, using area-weighted sampling to ensure higher density in regions with greater surface complexity. This DSM-guided geometric prior is essential for physically accurate shadow simulation and stable downstream optimization.

Multi-scale Geometric Optimization with Monocular Depth Supervision

The pipeline augments SatSplat [satsplat] by introducing:

• Monocular depth loss (based on a foundation monocular depth model [depth_anything_v2]) applied at early training stages to accelerate convergence and regularize optimization in regions with weak stereo overlap.
• Multi-scale geometric refinement: Affine camera models with randomized orientations and zoom factors synthesize supervisory images at variable resolutions, enabling Gaussians to learn robust geometry even for high-rise or sub-pixel-level facade details where direct satellite-sensor coverage is sparse.
• Adaptive Gaussian densification: Scale-based densification and improved opacity regularization encourage solid surface reconstruction and suppress transparent artifacts or disconnected primitives.

Shadow-Guided Generative Refinement

Key to SatSplatDiff is a shadow-guided generative refinement process:

• Novel view synthesis uses affine camera sampling to render novel geometries and shadow simulations informed by physical sun angles, providing illumination-diverse inputs for downstream refinement.
• Shadow-conditioned diffusion: Existing image-to-image diffusion models (e.g., FLUX.2 [flux-2-2025]) are conditioned on shadow-cast rendered images, using carefully engineered prompts to encourage ultra-sharp, rectilinear, sensor-consistent outputs. The diffusion network refines appearance while preserving pre-computed geometric shadow boundaries, thus inhibiting hallucinated geometry changes.
• Mixed supervision: The GS optimization is supervised on a pseudo-dataset mixing original and diffusion-enhanced images, balancing real data statistics with hallucination-free photorealism.

Losses and Optimization

Losses span photometric reconstruction ($\mathcal{L}_1$ + SSIM), geometric regularization (distortion, normal, shadow entropy, opacity entropy), monocular depth correlation, and distributional alignment (FID-CLIP [kynkaanniemi2022role], CMMD [jayasumana2024rethinking], LPIPS [zhang2018unreasonable]). Joint training in stages, with careful learning rate and densification scheduling, delivers stable, scalable optimization even under diverse acquisition geometries and radiometries.

Experimental Analysis

Evaluation on DFC2019 and IARPA2016 benchmarks demonstrates that SatSplatDiff consistently outperforms classical stereo, direct NeRF, GS, and generative refinement baselines (e.g., EOGS [aira2025gaussian], Skyfall-GS [lee2025SkyfallGS], Sat-NGP [billouard2024sat]) both qualitatively and across all numeric metrics:

• Geometric accuracy: SatSplatDiff reduces $\mathrm{MAE}_\mathrm{reg}$ (registration-based mean absolute error to LiDAR) by up to 18% (mean 1.23 m over full scenes), with consistent sub-meter accuracy on buildings and strong suppression of floating and disconnected artifacts.
• Visual fidelity: Distributional distance to real imagery (FID-CLIP) improves by 28–45% compared to the most competitive baselines; image structure and perceptual similarity (CW-SSIM, LPIPS) also show marked gains, especially for facades and edges.
• Resolution and cross-tile consistency: The pipeline supports up to $5\times$ effective resolution enhancement, and reconstructions from adjacent tiles are directly mappable with minimal seams or radiometric drift.

Ablation studies confirm the necessity of each major innovation:

• Removing multi-scale geometric supervision or monocular depth signals yields higher error and more visually inconsistent surfaces.
• Eliminating shadow-guided conditioning in generative refinement produces hallucinated structures and significant geometric degradation.
• Excessive diffusion-generated pseudo-views (>180 per tile) slightly impair performance due to distributional drift, indicating an optimal regime for supervision balancing.

Practical and Theoretical Implications

SatSplatDiff demonstrates that geometry-preserving generative refinement is achievable without sacrificing physical accuracy, provided geometric cues—particularly shadow boundaries—are propagated throughout both optimization and refinement. This methodology addresses critical gaps for remote sensing, surveillance, and mapping applications where both geometric integrity and visual realism are fundamental, and where nadir-centric satellite views are the norm.

The integration of multi-modal supervision democratizes high-fidelity 3D urban mapping from commercial satellite constellations, with practical scalability to city-scale deployment. Requirements for robust initial DSMs, accurate metadata, and sufficient view redundancy (≥12 images per tile) remain, though these align with established photogrammetric best practices.

On the theoretical front, the work validates that deep diffusion priors can be harnessed for 3D generative refinement in a controllable, geometry-consistent fashion—contradicting widely-held concerns about inevitable hallucination and geometric corruption in such settings. The approach also highlights new directions for neural rendering: explicit geometric constraint propagation, domain-aware augmentation, and the blending of model-based and model-free pipelines.

Future Directions

Open directions include full automation of DSM bootstrapping under limited coverage, incorporation of semantic or material priors for urban parsing, further robustness to sensor or illumination outliers, and truly end-to-end neural optimization with tightly coupled photometric, geometric, and generative cues. Cross-domain generalization (e.g., from satellite to aerial or ground-level imagery) may benefit from similar hybrid approaches. Accelerating diffusion-based refinement for real-time or global-scale applications is also an impactful avenue given computational overheads.

Conclusion

SatSplatDiff (2606.27223) achieves state-of-the-art results for high-fidelity, geometry-consistent 3D reconstruction from multi-date satellite imagery. The fusion of physically grounded shadow constraints and generative refinement via diffusion models enables texture enhancement without geometric compromise, supported by systematic architectural and algorithmic innovations. The demonstrated gains across large-scale, complex urban benchmarks provide a robust path towards scalable, reliable 3D mapping in remote sensing and beyond.