FineSplat: Fine Patches, Local Groups, Detailed Splats

Hyebin Kim; Hwasik Jeong; Joonhyuk Park; Gyeongjin Kang; Jungwoo Kim; Younghyun Kim; Eunbyung Park

cs.CVOct 18, 2026

FineSplat: Fine Patches, Local Groups, Detailed Splats

Authors: Hyebin Kim, Hwasik Jeong, Joonhyuk Park, Gyeongjin Kang, Jungwoo Kim, Younghyun Kim, Eunbyung Park

Organizations: Yonsei University

Abstract

Feed-forward 3D Gaussian Splatting (3DGS) promises instant scene reconstruction with high-quality novel-view synthesis. Yet, fine-grained geometry and appearance remain difficult to recover, motivating us to revisit how feed-forward models represent, aggregate, and refine detailed visual information. We introduce FineSplat, a feed-forward 3DGS framework built around fine-grained image patches. Since global multi-view attention becomes prohibitively expensive at such high token resolutions, we develop a dual-grouping strategy that combines two complementary forms of locality: view-wise grouping enables efficient interaction among neighboring views, while 3D spatial grouping reconnects geometrically related tokens that may never have interacted under view-wise grouping by lifting them into the 3D scene space. We further leverage reconstruction error as direct feedback, providing an explicit rendering-based signal to recover fine-grained structures and appearance, thereby overcoming the limitations of one-pass feed-forward prediction. Finally, a compact Gaussian generation scheme preserves the benefits of fine-grained features without unnecessarily increasing the number of primitives. Together, these designs enable FineSplat to recover richer geometry and appearance while remaining efficient and compact, achieving state-of-the-art results by a substantial margin, under the same or smaller Gaussian budgets than existing methods.

Figures & tables

Figure 1: Left : Number of Gaussians and average PSNR in the pose-free setting on DL3DV (224 $\times$ 224 resolution, 24 context views). Right : Qualitative comparison of novel view rendering in the pose-free setting with 24 context views. Our FineSplat represents scenes more accurately and preserves the fine details while using the same or smaller Gaussian budgets.

Figure 2: Method Overview. FineSplat first performs view-wise grouping and applies self-attention within each group, enabling efficient processing of fine-grained tokens. 3D spatial grouping lifts tokens into 3D space and enables interactions among spatially neighboring tokens, including those from different view groups, to enhance cross-view information exchange. Compact Gaussian generation predicts fewer Gaussians per token than existing pixel-aligned methods, maintaining a compact 3DGS representation. Finally, direct error feedback incorporates reconstruction errors obtained by rendering the input views to further refine token representations.

Figure 6: Neighbors within a serialized attention window in 3D space. Tokens in the same 3D spatial window share a color. Star : The query tokens in its source view. Circle : 3D positions of the tokens in the attention window projected onto the query’s source view. Square : Image patches of the tokens in the attention window in their own source views.

Step	6 views			12 views			24 views
Step	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
0	26.109	0.852	0.133	25.794	0.846	0.141	25.776	0.850	0.141
1	26.959	0.866	0.121	26.741	0.862	0.127	26.814	0.867	0.125
2	27.201	0.870	0.117	27.005	0.866	0.123	27.097	0.872	0.121
3	27.289	0.871	0.116	27.100	0.867	0.121	27.193	0.873	0.119
4	27.325	0.871	0.115	27.140	0.868	0.121	27.233	0.873	0.119

Table 4: Rendering quality at each Stage 2 step on DL3DV with 6, 12, and 24 input views. Step 0 denotes the prediction from Stage 1. The model is trained with four steps of Stage 2, and gray rows go beyond this training setting.

Figure 7: PSNR against reconstruction time on DL3DV in the posed setting, with each marker annotated by the number of context views.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: View selection for $e^{\text{agg}}$ . The query token in the source view (center) is projected into the other context views. Star : query token in its source view. Square : projected location in a view that passes both the frustum and depth tests and is included in $e^{\text{agg}}$ . Cross : projected location in a view that fails the depth test. Red-tinted view : view that fails the frustum test.

Figure 9: Comparison of resizing filters. From left to right, an original high-resolution image and its downsampled versions using PIL Lanczos and bilinear interpolation with anti-aliasing (Bilinear+AA). Lanczos better preserves high-frequency details such as the thin cracks and surface texture, whereas Bilinear+AA produces a smoother image in which these details are partly lost.

Method	$p$	$k$	6 views				12 views				24 views
Method	$p$	$k$	# GS	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	# GS	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	# GS	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
DepthSplat	✓	✓	393K	24.63	0.837	0.142	786K	22.64	0.793	0.187	1.6M	20.60	0.734	0.248
DA3	✓	✓	1.5M	23.99	0.805	0.158	3.0M	22.84	0.758	0.190	6.1M	21.70	0.710	0.230
ZipSplat	✓	✓	62K	24.75	0.790	0.174	124K	23.40	0.735	0.204	249K	23.20	0.732	0.206
Ours	✓	✓	95K	28.33	0.886	0.111	191K	28.15	0.883	0.117	381K	28.25	0.889	0.116
Ours prune	✓	✓	62K	27.33	0.868	0.138	124K	27.31	0.867	0.141	249K	27.64	0.876	0.134

Appendix

Table 5: Novel view synthesis results on DL3DV with 6, 12, and 24 input views at

252\times 252

resolution, evaluated on images downsampled with bilinear interpolation with anti-aliasing (Bilinear+AA).

Stage	View-group size	6 views			12 views			24 views
Stage	View-group size	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
	1	21.928	0.780	0.183	23.165	0.800	0.172	23.851	0.817	0.163
	2	26.109	0.852	0.133	25.794	0.846	0.141	25.776	0.850	0.141
	4	–	–	–	25.456	0.839	0.147	25.447	0.844	0.146
	6	25.491	0.834	0.147	25.074	0.829	0.155	25.001	0.835	0.153
	8	–	–	–	–	–	–	24.628	0.827	0.160

Appendix

Table 6: Ablation of the view-group size on DL3DV with 6, 12, and 24 input views. The model is trained with a group size of 2, and only the group size of the view-wise grouping is changed at inference. Stage 1 denotes the initial prediction, and Stage 2 the output of the full model.

$e^{\text{self}}$	$e^{\text{agg}}$	6 views			12 views			24 views
$e^{\text{self}}$	$e^{\text{agg}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
	✓	26.890	0.864	0.121	26.586	0.858	0.129	26.611	0.863	0.127
✓		27.236	0.870	0.116	27.000	0.865	0.123	27.063	0.870	0.122
✓	✓	27.325	0.871	0.115	27.140	0.868	0.121	27.233	0.873	0.119

Appendix

Table 7: Ablation on the error feature on DL3DV with 6, 12, and 24 input views.

$e^{\text{agg}}$ views	6 views			12 views			24 views
$e^{\text{agg}}$ views	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
All (Ours)	27.325	0.871	0.115	27.140	0.868	0.121	27.233	0.873	0.119
Within Group	27.099	0.866	0.118	26.734	0.859	0.128	26.628	0.860	0.131
Random	27.141	0.867	0.118	26.816	0.860	0.126	26.801	0.863	0.128

Appendix

Table 8: Masking out the views outside the view group in

e^{\text{agg}}

at inference time. Within group :

e^{\text{agg}}

averages only the views of the token’s own view group in Stage 1. Random : the same number of visible views is removed at random.

Figure 11: Qualitative comparison on DL3DV. Rows 1 to 4, 5 to 8, and 9 to 12 use 6, 12, and 24 context views, respectively.

Figure 12: Qualitative comparison on RE10K with 6 input views. Our model is evaluated zero-shot without training on RE10K, whereas ZipSplat is trained on RE10K.

Figure 13: Qualitative results of cross-dataset generalization from DL3DV to ScanNet++ (the first six columns) and Mip-NeRF 360 (the last three columns). Rows 1 to 3, 4 to 6, and 7 to 9 correspond to 32, 64, and 128 context views, respectively.

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,013 papers

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens
Apr 16, 2026Roni Itkin, Noam Issachar, Yehonatan Keypur +33D Gaussian Splatting 3DGS Acceleration
The Hebrew University of Jerusalem · Westlake University
ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion
Jul 22, 2026In Cho, Jeonghwan Cho, Mijin Yoo +23D Gaussian Splatting Adaptive Gaussian Densification
Yonsei University, South Korea · National University of Singapore, Singapore
ZipSplat: Fewer Gaussians, Better Splats
Jun 3, 2026Alexander Veicht, Sunghwan Hong, Dániel Baráth +13D Gaussian Splatting Multi-View 3D Reconstruction
ETH Zürich · Microsoft

Apr 16, 2026J/K move · Enter open · S save

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

Roni Itkin, Noam Issachar, Yehonatan Keypur, Xingyu Chen+2

The Hebrew University of Jerusalem · Westlake University

3D Gaussian Splatting 3DGS Acceleration Novel View Synthesis 3D Scene Representation Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

The efficient spatial allocation of primitives serves as the foundation of 3D Gaussian Splatting, as it directly dictates the synergy between representation compactness, reconstruction speed, and rendering fidelity. Previous solutions, whether based on iterative optimization or feed-forward inference, suffer from significant trade-offs between these goals, mainly due to the reliance on local, heuristic-driven allocation strategies that lack global scene awareness. Specifically, current feed-forward methods are largely pixel-aligned or voxel-aligned. By unprojecting pixels into dense, view-aligned primitives, they bake redundancy into the 3D asset. As more input views are added, the representation size increases and global consistency becomes fragile. To this end, we introduce GlobalSplat, a framework built on the principle of align first, decode later. Our approach learns a compact, global, latent scene representation that encodes multi-view input and resolves cross-view correspondences before decoding any explicit 3D geometry. Crucially, this formulation enables compact, globally consistent reconstructions without relying on pretrained pixel-prediction backbones or reusing latent features from dense baselines. Utilizing a coarse-to-fine training curriculum that gradually increases decoded capacity, GlobalSplat natively prevents representation bloat. On RealEstate10K and ACID, our model achieves competitive novel-view synthesis performance while utilizing as few as 16K Gaussians, significantly less than required by dense pipelines, obtaining a light 4MB footprint. Further, GlobalSplat enables significantly faster inference than the baselines, operating under 78 milliseconds in a single forward pass. Project page is available at https://r-itk.github.io/globalsplat/

Abstract

3D Gaussian Splatting

Explore similar work

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion

ZipSplat: Fewer Gaussians, Better Splats