ProDiG: Progressive Diffusion-Guided Gaussian Splatting for Aerial to Ground Reconstruction

Sirshapan Mitra; Yogesh S. Rawat

cs.CVApr 2, 2026

ProDiG: Progressive Diffusion-Guided Gaussian Splatting for Aerial to Ground Reconstruction

Authors: Sirshapan Mitra, Yogesh S. Rawat

Organizations: CRCV, University of Central Florida

Abstract

Generating ground-level views and coherent 3D site models from aerial-only imagery is challenging due to extreme viewpoint changes, missing intermediate observations, and large scale variations. Existing methods either refine renderings post-hoc, often producing geometrically inconsistent results, or rely on multi-altitude ground-truth, which is rarely available. Gaussian Splatting and diffusion-based refinements improve fidelity under small variations but fail under wide aerial-toground gaps. To address these limitations, we introduce ProDiG (Progressive Diffusion-Guided Gaussian Splatting for Aerial to Ground Reconstruction), a diffusionguided framework that progressively transforms aerial 3D representations toward ground-level fidelity. ProDiG synthesizes intermediate-altitude views and refines the Gaussian representation at each stage using a geometry-aware causal attention module that injects epipolar structure into reference-view diffusion. A distance-adaptive Gaussian module dynamically adjusts Gaussian scale and opacity based on camera distance, ensuring stable reconstruction across large viewpoint gaps. Together, these components enable progressive, geometrically grounded refinement without requiring additional ground-truth viewpoints. Extensive experiments on synthetic and real-world datasets demonstrate that ProDiG produces visually realistic ground-level renderings and coherent 3D geometry, significantly outperforming existing approaches in terms of visual quality, geometric consistency, and robustness to extreme viewpoint changes. Project Page: https://sirsh07.github.io/research/prodig

Figures & tables

Figure 1: Overview of ProDiG. (a) Our framework reconstructs a complete 3D scene using only aerial images. A large distribution shift exists between the aerial training images and the ground-level query images. During evaluation, we render novel views at ground-level camera poses and compare them against ground-truth images. (b) In our Distance-Adaptive Gaussian Splatting module, each Gaussian is dynamically scaled and reweighted using a lightweight encoder that predicts adjustment factors from its learned scaling feature and its distance to the active camera. (c) We progressively render noisy novel views at successively lower altitudes, fix these views using our diffusion model, and iteratively retrain the Gaussian Splatting model using the fixed novel view.

Figure 2: Overview of aeroFix: (left) Our diffusion model is fine-tuned on aerial imagery using LoRA. The noisy novel view is fixed using the reference view to fixed novel image. In the diffusion block, the relative camera pose difference is injected into the timestep embedding of the noisy image to encode geometric variation across viewpoints. We additionally include Plücker ray embeddings before the attention mixing layer to provide geometric cues. In the Causal Attention Mixing module, we enforce an epipolar constraint by masking the novel query - reference key attention map such that only tokens aligned with the corresponding epipolar lines retain attention (value 1), while all others are suppressed (value 0). The reference query - novel key block is fully masked, and the remaining attention blocks operate under standard full attention. (right) The figure illustrates epipolar correspondences for multiple query points on the noisy image and their corresponding lines on the reference view.

Figure 3: Effectiveness of aeroFix: Comparison of aerial image refinement between Difix3D+ [ 38 ] and our aeroFix model. The noisy novel views are outlined in orange , the reference images in green , and the refined (fixed) novel images in pink . Difix3D+ tends to copy content from the reference view when the viewpoint difference is large, leading to inconsistencies and artifacts. In contrast, aeroFix effectively preserves structural fidelity and produces geometrically consistent, artifact-free refinements.

Figure 4: Qualitative analysis of ProDiG(ours): Comparison of our method with existing baselines on aerial-to-ground reconstruction. Gaussian Splatting [ 16 ] struggles to render complete scenes due to the absence of ground-level viewpoints, while Difix3D+ [ 38 ] exhibits noisy artifacts and hallucinated structures. In contrast, ProDiG (ours) produces geometrically consistent and visually coherent reconstructions with fewer hallucinations. Notably, in the second row, the aerial inputs and reconstructed model include cars visible from above, whereas the ground-truth image - captured at a different time - does not.

Method	D $\downarrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$
Difix3D+ [ 38 ]	0.15	20.47	0.54	0.42
Difix3D+ (LoRA)	0.07	21.45	0.59	0.30
Pose + Plücker	0.06	22.30	0.64	0.27
Pose + Plücker + Causal	0.03	23.35	0.68	0.24
aeroFix (Ours)	0.03	23.68	0.69	0.24

Table 1: Comparison of aeroFix with Difix3D+ [ 38 ] . Best values are in bold .(D: Dreamsim, P: PSNR, S: SSIM, L: LPIPS)

Method	Wriva S06				Wriva S01
	D $\downarrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$	D $\downarrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$
3DGS [ 16 ]	0.79	7.74	0.27	0.94	0.66	9.74	0.40	0.85
3DGS-MCMC [ 18 ]	0.71	8.71	0.31	0.90	0.57	9.73	0.39	0.80
2DGS [ 12 ]	0.78	7.08	0.21	0.95	0.68	9.68	0.36	0.85
Scaffold-GS [ 24 ]	0.72	7.79	0.22	0.88	0.67	10.35	0.37	0.82
Difix3D+ [ 38 ]	0.66	8.91	0.28	0.80	0.38	11.83	0.37	0.67

Table 2: Comparison of our method with existing methods on the WRIVA dataset across two sites. Best values are in bold .(D: Dreamsim, P: PSNR, S: SSIM, L: LPIPS)

Method	D $\downarrow$	P $\uparrow$	S $\uparrow$	L $\downarrow$
3DGS [ 16 ]	0.51	10.71	0.40	0.77
2DGS [ 12 ]	0.62	9.29	0.28	0.81
3DGS-MCMC [ 18 ]	0.49	10.84	0.41	0.77
Scaffold-GS [ 24 ]	0.54	10.19	0.35	0.75
Difix3D+ [ 38 ]	0.48	11.38	0.38	0.63
Ours	0.39	12.39	0.41	0.50

Table 3: Comparison of our method with existing methods on the MatrixCity dataset. Best values are in bold . (D: Dreamsim, P: PSNR, S: SSIM, L: LPIPS)

Figure 5: Generalization across Varying Altitudes. We evaluate our method on the Aerial MegaDepth [ 36 ] dataset, which contains sites captured at diverse altitude ranges.

Figure 6: Ablations. (top) Comparison of different progressive methods. (bottom) Effectiveness of Distance Adaptive Gaussian Module. 7k represents, evalution at 7k iteration, after initial training.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 930 papers

Feed-Forward Gaussian Splatting from Sparse Aerial Views
May 19, 2026Dongli Wu, Zhuoxiao Li, Tongyan Hua +43D Gaussian Splatting 3D Reconstruction
1The Hong Kong University of Science and Technology (Guangzhou) · 2Peking University · 3The Ohio State University
D3GS: Depth, DINO, and RGB Diffusion Co-Guided 3D Gaussian Splatting for Sparse-View Reconstruction
Sep 19, 2026Yunqi Gao, Zhanfeng Liao, Hanzhang Tu +7Sparse-View 3D Reconstruction 3D Gaussian Splatting
Central China Normal University · Tsinghua University · Fuzhou University +4
ReconSplat: Generalizable 3D Scene Reconstruction Beyond Observed Views
Aug 28, 2026Giuseppe Stracquadanio, Kevin Raj, Julia Grabinski +1Feed-Forward 3D Reconstruction Novel View Synthesis
TU Darmstadt · Zuse School ELIZA

May 19, 2026J/K move · Enter open · S save

Feed-Forward Gaussian Splatting from Sparse Aerial Views

Dongli Wu, Zhuoxiao Li, Tongyan Hua, Yinrui Ren+3

1The Hong Kong University of Science and Technology (Guangzhou) · 2Peking University · 3The Ohio State University

3D Gaussian Splatting 3D Reconstruction Sparse-View 3D Reconstruction Urban Scene Reconstruction Feed-Forward 3D Reconstruction Aerial 3D Reconstruction

Open paper AskPDF arXiv

Reconstructing large-scale urban scenes from sparse aerial views is a crucial yet challenging task. Due to biased top-down and shallow-oblique camera poses, sparse aerial captures exhibit strong evidence imbalance: roofs and open regions are repeatedly observed, while facades, distant buildings, and occluded structures receive little multi-view support. Existing feed-forward 3D Gaussian Splatting methods directly regress a deterministic representation from sparse inputs, but this often leads to ghosting, melted facades, and stretched textures. Recent pseudo-view and video-based generative reconstruction methods use additional supervision or generative priors. However, they often lack a clear separation between observed geometry and prior-driven content, which can lead to plausible but inconsistent structures. We propose AnyCity, an observation-grounded generative reconstruction framework for sparse aerial urban scenes. AnyCity first predicts an observation-supported geometry latent to anchor reliable structures, and then uses scaffold-conditioned aerial completion tokens to predict a gated residual update for weakly constrained content before Gaussian decoding. During training, dense-to-sparse distillation transfers structural cues from dense-view reconstruction, while an aerial-adapted video diffusion prior provides fine-grained urban appearance cues through gated token conditioning. Observation-preserving objectives keep the refined representation consistent with input-supported geometry. At inference time, AnyCity reconstructs the final 3D Gaussian scene from sparse aerial views in a single feed-forward pass, achieving coherent urban novel-view synthesis with second-level inference. Experiments on synthetic, aerial-domain, UAV-textured, and real-world scenes show consistent improvements over feed-forward baselines.

Abstract

3D Reconstruction

Explore similar work

Feed-Forward Gaussian Splatting from Sparse Aerial Views

D3GS: Depth, DINO, and RGB Diffusion Co-Guided 3D Gaussian Splatting for Sparse-View Reconstruction

ReconSplat: Generalizable 3D Scene Reconstruction Beyond Observed Views