PixGR: Pixel-Space Generative Decoding for Fast Feed-Forward 3D Reconstruction

Ranran Huang; Wenjing Bian; Weixun Luo; Ye Mao; Krystian Mikolajczyk

cs.CVOct 22, 2026

PixGR: Pixel-Space Generative Decoding for Fast Feed-Forward 3D Reconstruction

Authors: Ranran Huang, Wenjing Bian, Weixun Luo, Ye Mao, Krystian Mikolajczyk

Organizations: Imperial College London · University of Tübingen

Abstract

We present PixGR, an efficient framework that integrates lightweight pixel-space generative modeling into feed-forward 3D Gaussian reconstruction to model scene uncertainty. Context observations are encoded once into a scene latent representation, which conditions generative decoding from noisy target views to produce a shared 3D Gaussian representation. We formulate this process using pixel-space flow matching, enabling end-to-end optimization of 3D reconstruction with direct pixel-space generative supervision through differentiable rendering. This avoids VAE-based latent compression and error propagation inherent in previous two-stage pipelines. To preserve feed-forward efficiency, PixGR caches the scene representation and performs iterative inference using a lightweight DiT decoder. Experiments demonstrate that PixGR outperforms regression-based approaches in challenging sparse-view settings, particularly with widely spaced observations and view extrapolation. Meanwhile, it maintains sub-second inference and remains competitive with substantially heavier pretrained generative models. Our project page is available at https://ranrhuang.github.io/pixgr/.

Figures & tables

Figure 1: Comparison of feed-forward reconstruction and generative pipelines. (a) Deterministic reconstruction cannot model scene uncertainty. Existing generative approaches address this limitation through (b) generation then reconstruction or (c) reconstruction then generation, but suffer from error propagation between stages, VAE-based information loss, and high computational costs. (d) PixG3R instead integrates lightweight pixel-space generation directly into 3D Gaussian reconstruction, enabling end-to-end optimization with direct pixel-space generative supervision.

Figure 2: PixGR Overview. Multi-view posed context images are encoded once into a scene latent representation. Target viewpoints are sampled and encoded as ray maps, which are combined with noisy target images. A lightweight generative decoder processes the concatenated scene latent and target tokens to predict a shared set of pixel-aligned 3D Gaussians supporting both context and target views. The decoder adopts a DiT-style architecture with global self-attention, followed by a linear Gaussian prediction head. Differentiable rendering of the predicted Gaussians produces denoised target views, enabling direct supervision through pixel-space generative objectives.

Regression-based Methods
Method	RE10K				DL3DV
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
MVSplat	26.39	0.869	0.128	2.28	18.12	0.556	0.370	58.20
DepthSplat	27.47	0.889	0.114	2.29	24.19	0.823	0.147	9.51
AnySplat	19.80	0.682	0.274	2.88	19.03	0.554	0.235	16.94
YoNoSplat	25.73	0.851	0.130	3.75	24.83	0.821	0.137	11.56

Table 1: Quantitative comparisons of 3D reconstruction performance on view interpolation. Results are reported on 2-view RE10K and 6-view DL3DV. PixGR consistently outperforms prior methods across datasets. Best and second-best results are highlighted.

Figure 3: Qualitative comparison on interpolated views. PixGR produces sharper and more geometrically consistent renderings than prior methods and handles large viewpoint changes effectively.

Regression-based Methods
Method	RE10K				DL3DV
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
DepthSplat	22.02	0.822	0.189	6.95	21.59	0.745	0.205	17.41
AnySplat	16.21	0.621	0.351	5.98	14.84	0.449	0.426	22.16
YoNoSplat	23.00	0.810	0.182	6.47	23.04	0.761	0.192	17.21
Generative Methods

Table 2: Quantitative comparisons of 3D reconstruction performance on view extrapolation. Results are reported on 2-view RE10K and 4-view DL3DV. PixGR consistently outperforms prior methods.

Figure 4: Qualitative comparison on extrapolated novel views. Regression-based methods fail in unseen regions, while generative baselines may produce artifacts and distortions, and lead to geometric misalignment (e.g., the sofa in Row 1) and detail loss (e.g., missing texture on the dustbin and the missing street lamp in Row 3). In contrast, PixGR produces geometrically consistent renderings while preserving fine details.

Figure 5: Inference time versus the number of novel views on a single A100 GPU.

Encoder ablation
Variant	Interpolation				Extrapolation
Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
VGGT w/o init.	27.32	0.882	0.134	2.45	24.98	0.859	0.167	3.67
ViT-L encoder	27.14	0.881	0.126	2.63	24.80	0.848	0.170	3.89
Decoder ablation
Regressor	27.58	0.883	0.111	2.38	24.34	0.821	0.166	5.42

Table 3: Ablation of encoder and decoder design on RE10K under interpolation and extrapolation. PixGR uses a pretrained VGGT encoder and generative decoder, achieving the best performance.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Regression-based Methods
Method	Prior		Params (B)
Method	Geometric	Generative	Encoder	Decoder	Total
DepthSplat	DA2 ( Yang et al., 2024 )	–	0.347	–	0.347
AnySplat	VGGT ( Wang et al., 2025 )	–	1.191	–	1.191
YoNoSplat	Pi3 ( Wang et al., 2026 )	–	0.965	–	0.965
DA3	DA3 ( Lin et al., 2025b )	–	1.356	–	1.356

Appendix

Table 4: Geometric and generative priors and parameter counts of the compared methods. A dash indicates that no corresponding prior is used or that a separate parameter count is unavailable.

Regression-based Methods
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
MVSplat	28.25	0.843	0.144	6.38
DepthSplat	19.65	0.628	0.298	28.36
AnySplat	21.49	0.631	0.300	8.47
YoNoSplat	25.72	0.768	0.185	10.14
Generative Methods

Appendix

Table 5: Out-of-domain novel-view synthesis results on ACID.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
Depth Anything V3	17.91	0.574	0.339	13.28
NoPoSplat	21.41	0.757	0.207	6.09
NAS3R	19.92	0.782	0.205	7.78
SPFSplat	20.22	0.775	0.210	7.49
PixGR	26.03	0.872	0.141	2.75

Appendix

Table 6: Comparison with additional regression-based methods on 2-view RE10K extrapolation.

Feed-Forward 3D Reconstruction

Feed-forward 3D reconstruction predicts geometry, cameras, and renderable scene representations directly from images without per-scene optimization.

See all 313 papers

GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views
Feb 26, 2026Tianyu Chen, Wei Xiang, Kang Han +43D Reconstruction Sparse-View 3D Reconstruction
La Trobe University, Melbourne, VIC 3086, Australia · Cisco Research, San Jose, CA, USA
PixWorld: Unifying 3D Scene Generation and Reconstruction in Pixel Space
Jul 6, 2026Sensen Gao, Zhaoqing Wang, Qihang Cao +33D Reconstruction Pixel-Space Diffusion Models
Nanyang Technological University · AISphere
Free-Range Gaussians: Non-Grid-Aligned Generative 3D Gaussian Reconstruction
Apr 6, 2026Ahan Shabanov, Peter Hedman, Ethan Weber +103D Reconstruction Multi-View 3D Reconstruction
Meta Reality Labs · Simon Fraser University · University of Oxford +1

Feb 26, 2026J/K move · Enter open · S save

GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views

Tianyu Chen, Wei Xiang, Kang Han, Yu Lu+3

La Trobe University, Melbourne, VIC 3086, Australia · Cisco Research, San Jose, CA, USA

3D Reconstruction Sparse-View 3D Reconstruction Feed-Forward 3D Reconstruction 3D Gaussian Splatting

Open paper AskPDF arXiv

Feed-forward 3D reconstruction offers substantial runtime advantages over per-scene optimization, which remains slow at inference and often fragile under sparse views. However, existing feed-forward methods still have potential for further performance gains, especially for out-of-domain data, and struggle to retain second-level inference time once a generative prior is introduced. These limitations stem from the one-shot prediction paradigm in existing feed-forward pipeline: models are strictly bounded by capacity, lack inference-time refinement, and are ill-suited for continuously injecting generative priors. We introduce GIFSplat, a purely feed-forward iterative refinement framework for 3D Gaussian Splatting from sparse unposed views. A small number of forward-only residual updates progressively refine current 3D scene using rendering evidence, achieve favorable balance between efficiency and quality. Furthermore, we distill a frozen diffusion prior into Gaussian-level cues from enhanced novel renderings without gradient backpropagation or ever-increasing view-set expansion, thereby enabling per-scene adaptation with generative prior while preserving feed-forward efficiency. Across DL3DV, RealEstate10K, and DTU, GIFSplat consistently outperforms state-of-the-art feed-forward baselines, improving PSNR by up to +2.1 dB, and it maintains second-scale inference time without requiring camera poses or any test-time gradient optimization.

Abstract

Feed-Forward 3D Reconstruction

Explore similar work

GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views

PixWorld: Unifying 3D Scene Generation and Reconstruction in Pixel Space

Free-Range Gaussians: Non-Grid-Aligned Generative 3D Gaussian Reconstruction