Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting

Yixing Lao; Xuyang Bai; Xiaoyang Wu; Nuoyuan Yan; Zixin Luo; Tian Fang; Jean-Daniel Nahmias; Yanghai Tsin

cs.CVMar 26, 2026

Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting

Authors: Yixing Lao, Xuyang Bai, Xiaoyang Wu, Nuoyuan Yan, Zixin Luo, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, +2 more

Organizations: HKU · Apple

Abstract

Existing feed-forward 3D Gaussian Splatting methods predict pixel-aligned primitives, leading to a quadratic growth in primitive count as resolution increases. This fundamentally limits their scalability, making high-resolution synthesis such as 4K intractable. We introduce LGTM (Less Gaussians, Texture More), a feed-forward framework that overcomes this resolution scaling barrier. By predicting compact Gaussian primitives coupled with per-primitive textures, LGTM decouples geometric complexity from rendering resolution. This approach enables high-fidelity 4K novel view synthesis without per-scene optimization, a capability previously out of reach for feed-forward methods, all while using significantly fewer Gaussian primitives. Project page: https://yxlao.github.io/lgtm/

Figures & tables

Rendering Resolution	Method	Primitive Resolution	Texture Size	Metrics			Train Memory ( $\text{bs}=1$ )
Rendering Resolution	Method	Primitive Resolution	Texture Size	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	Train Memory ( $\text{bs}=1$ )
1024 $\times$ 576	NoPoSplat	1024 $\times$ 576	-	0.239	0.716	23.169	61.85 GB
	LGTM	512 $\times$ 288	2 $\times$ 2	0.213	0.816	25.606	20.16 GB
	LGTM	256 $\times$ 144	4 $\times$ 4	0.283	0.762	24.135	16.26 GB
2048 $\times$ 1152	NoPoSplat	2048 $\times$ 1152	-	✗	✗	✗	OOM
	LGTM	512 $\times$ 288	4 $\times$ 4	0.176	0.810	25.328	21.39 GB

Table 1: LGTM enables high-resolution feed-forward Gaussian splatting. We compare LGTM with NoPoSplat Ye et al. (2025) variants using different primitive resolutions but yielding the same effective output resolution. NoPoSplat fails to train at 2K or higher due to memory and compute limits, whereas LGTM reaches equivalent resolutions with compact geometric primitives and per-primitive texture maps. Training memory reports peak GPU usage with batch size 1, 2 context views, and 4 target views. For inference performance and memory analysis, see Table 4 .

Figure 2: LGTM Architecture. Top : The primitive network $f_{\text{prim}}$ takes low-resolution images as input and predicts compact geometric primitives ${\bm{\mu}}$ , ${\bm{s}}$ , ${\bm{r}}$ , ${\bm{c}}$ . Bottom : The texture network $f_{\text{texture}}$ processes high-resolution images through image patchify and projective mapping networks, and predicts per-primitive texture maps ${\bm{T}}^{\alpha},{\bm{T}}^{c}$ . This decoupling of geometry and appearance enables LGTM to achieve feed-forward 4K Gaussian splatting with significantly fewer primitives.

Dataset	Method		Render Resolution	Primitive Resolution	Texture Size	Metrics
Dataset	Baseline	Primitive	Render Resolution	Primitive Resolution	Texture Size	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
RE10K	NoPoSplat	3DGS	1024 $\times$ 576 (1K)	512 $\times$ 288	-	0.250	0.823	24.666
		2DGS	1024 $\times$ 576 (1K)	512 $\times$ 288	-	0.225	0.829	24.674
		LGTM	1024 $\times$ 576 (1K)	512 $\times$ 288	2 $\times$ 2	0.195	0.847	25.419
	NoPoSplat	3DGS	2048 $\times$ 1152 (2K)	512 $\times$ 288	-	0.257	0.819	24.235
		2DGS	2048 $\times$ 1152 (2K)	512 $\times$ 288	-	0.227	0.823	24.282

Table 2: Novel view synthesis results under two-view setting. Performance of LGTM over NoPoSplat and DepthSplat across rendering resolutions and datasets. For 4K resolution, we use 4096

\times

2304 for NoPoSplat and 3840

\times

2048 for DepthSplat due to the requirements of their backbone networks.

Panel of Figure 3 — Figure 3: Qualitative comparison in the two-view setting. Evaluated on the DL3DV dataset at 4K resolution. Best viewed in color and zoomed in.

Inputs	Method		Render Resolution	Primitive Resolution	Texture Size	Metrics
Inputs	Baseline	Primitive	Render Resolution	Primitive Resolution	Texture Size	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
Single-view	Flash3D	3DGS	1024 $\times$ 576 (1K)	512 $\times$ 288	-	0.274	0.672	21.353
		2DGS	1024 $\times$ 576 (1K)	512 $\times$ 288	-	0.267	0.676	21.447
		LGTM	1024 $\times$ 576 (1K)	512 $\times$ 288	2 $\times$ 2	0.215	0.708	21.910
	Flash3D	3DGS	2048 $\times$ 1152 (2K)	512 $\times$ 288	-	0.368	0.643	19.911
		2DGS	2048 $\times$ 1152 (2K)	512 $\times$ 288	-	0.358	0.653	20.276

Table 3: Single-view and multi-view novel view synthesis results. Comparison of LGTM with Flash3D Szymanowicz et al. (2025) and VGGT Wang et al. (2025a) baselines across different view settings and resolutions on DL3DV dataset.

Panel of Figure 4 — Figure 4: Qualitative comparison in the single-view setting. We re-train Flash3D on DL3DV with the same number of geometric primitives (512 $\times$ 288) for 3DGS, 2DGS, and LGTM.

	Method	Primitive Resolution	Texture Size	Render Resolution	Peak GPU Memory (GB)	Network Fwd. Time (ms)	Render Time (ms)	Total Time (ms)
①	NoPoSplat 3DGS	512 $\times$ 288	-	512 $\times$ 288	3.06	110.37	3.81	114.18
②	NoPoSplat 2DGS	512 $\times$ 288	-	512 $\times$ 288	3.06	112.70	6.43	119.13
③	LGTM	512 $\times$ 288	2 $\times$ 2	1024 $\times$ 576	4.33	132.06	7.70	139.76
④	LGTM	512 $\times$ 288	4 $\times$ 4	2048 $\times$ 1152	4.60	136.14	14.26	150.40
⑤	LGTM	512 $\times$ 288	8 $\times$ 8	4096 $\times$ 2304	5.51	142.28	32.82	175.10

Table 4: Inference time and memory benchmark. LGTM is highly efficient when scaling up to high resolutions: compared to the NoPoSplat 512

\times

288 2DGS (②), the LGTM 4096

\times

2304 model (⑤) represents a 64

\times

increase in pixels, yet requires only 1.80

\times

the peak memory and 1.47

\times

the total time. For an analysis of training memory requirements, see Table 1 .

	Method	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
①	NoPoSplat (3DGS)	0.371	0.583	16.964
②	+ High-res retrain (2DGS)	0.256	0.731	23.502
③	+ Image patchify only	0.199	0.782	24.673
④	+ Texture color	0.189	0.806	25.314
⑤	+ Texture alpha (full model)	0.176	0.810	25.328

Table 5: Ablation study on LGTM components. We start from the NoPoSplat baseline and progressively integrate components of LGTM, evaluated on DL3DV at 2K resolution. All models operate on 512

\times

288 primitives. The baseline is trained on low-resolution and evaluated at 2K. Subsequent versions are trained with high-resolution supervision.

Panel of Figure 5 — Figure 5: Qualitative results of the ablation study. Each image shows the progressive improvement as components of LGTM are added to the NoPoSplat baseline. The full model produces renderings that are visibly closer to the ground truth.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6: Effects of high-resolution re-training.

Figure 7: Unbounded bilinear texture sampling. Our sampling method uses border clamping to handle out-of-bounds coordinates, allowing texture information to extend beyond the primitive boundary. This provides smoother transitions and better coverage compared to bounded sampling approaches that use zero-padding for out-of-bounds areas.

Panel of Figure 8 — Figure 8: Qualitative comparison across different input view gaps. Each row shows context frame gaps of 10, 20, 30, and 40 frames. LGTM consistently produces sharper details and fewer artifacts.

Context view gap	Method	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$
10	NoPoSplat	0.322	0.737	22.198
10	NoPoSplat + LGTM	0.200	0.803	24.489
20	NoPoSplat	0.366	0.722	20.732
20	NoPoSplat + LGTM	0.258	0.765	22.159
30	NoPoSplat	0.413	0.709	19.430
30	NoPoSplat + LGTM	0.323	0.730	20.123

Appendix

Table 6: Quantitative comparison across different input view gaps. LGTM consistently outperforms the baseline across all gaps (10, 20, 30, 40 frames), demonstrating robustness to larger viewpoint differences.

Panel of Figure 9 — Figure 9: Comparing LGTM to per-scene optimization. Top: Per-scene optimization overfits to context views with degraded performance on middle frames. DepthSplat + LGTM maintains stable quality across all frames. Bottom: Target frame #10. While both achieve similar sharpness in the center, LGTM produces sharper details with fewer artifacts toward the edges.

Method	Per-Scene Optimization	DepthSplat + LGTM (Ours)
Frames for COLMAP	{0, 1, 2, …, 20}	-
Context views (optimization)	{0, 20}	{0, 20}
Target views (evaluation)	{1, 2, 3, …, 19}	{1, 2, 3, …, 19}
PSNR $\uparrow$	21.75	27.99
SSIM $\uparrow$	0.78	0.88
LPIPS $\downarrow$	0.21	0.15

Appendix

Table 7: Comparison with per-scene optimization. DepthSplat + LGTM outperforms per-scene optimized 3DGS while being orders of magnitude faster.

Explore similar work

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
LiteTex-GS: Fast and Lightweight Texturing for Gaussian Splatting
Sep 20, 2026Zhiwei Li, Yijia Guo, Yishi Lu +43D Gaussian Splatting Novel View Synthesis
School of Mathematics and Computer Science, Nanchang University · State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University · Henan University +4
StreamSplat: Streaming Feed-Forward 3D Gaussian Splatting
Aug 3, 2026Changhao Song, Yuxuan Wang, Qibiao Li +23D Gaussian Splatting Streaming 3D Reconstruction
University of Science and Technology of China

Jul 22, 2026J/K move · Enter open · S save

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

In Cho, Jeonghwan Cho, Mijin Yoo, Gim Hee Lee+1

Yonsei University, South Korea · National University of Singapore, Singapore

3D Gaussian Splatting Adaptive Gaussian Densification Novel View Synthesis Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) achieves high-quality novel-view synthesis by optimizing freely placed primitives in 3D and adaptively densifying them in under-reconstructed regions. However, this scene-adaptive capacity allocation is largely lost in existing feed-forward 3DGS methods, which commonly regress Gaussians at input pixels and lift them along camera rays. Such pixel-aligned formulations make the number and placement of primitives depend on image resolution and input viewpoints rather than scene complexity, resulting in dense and often redundant Gaussian sets. We present ATSplat, a feed-forward 3DGS framework that restores the adaptive allocation capability of 3DGS optimization through Adaptive 3D Tokens. ATSplat first lifts coarse patch-level depth and camera cues into sparse 3D anchor tokens, forming a compact scaffold of the scene. Each token is then regressed into local Gaussians with learnable 3D offsets, decoupling primitive placement from input image grids. An Adaptive Token Expansion module predicts a token-level uncertainty score, supervised by rendering error maps, and selectively expands high-uncertainty tokens through learnable expansion layers. This sparse-to-adaptive formulation enables ATSplat to concentrate primitives in challenging regions while maintaining a compact representation. Experiments on two representative datasets, RealEstate10K and DL3DV, show that ATSplat achieves state-of-the-art rendering quality while reducing the number of Gaussians by more than

5.7\times

compared with dense feed-forward 3DGS methods. From 12 input images at

512 \times 960

resolution, ATSplat completes reconstruction in less than a second using a single commercial GPU, and renders high-quality novel views at 1136 FPS (

512 \times 960

) with only 311K Gaussians.

Abstract

Explore similar work

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

LiteTex-GS: Fast and Lightweight Texturing for Gaussian Splatting

StreamSplat: Streaming Feed-Forward 3D Gaussian Splatting