DensiTok: Making Feed-Forward 3D Gaussian Splatting See More Views Than It Is Given

Minhyeok Lee; Jungho Lee; Minseok Kang; Heeseung Choi; Ig-Jae Kim; Sangyoun Lee

cs.CVOct 6, 2026

DensiTok: Making Feed-Forward 3D Gaussian Splatting See More Views Than It Is Given

Authors: Minhyeok Lee, Jungho Lee, Minseok Kang, Heeseung Choi, Ig-Jae Kim, Sangyoun Lee

Organizations: Yonsei University · NAVER AI Lab · Korea Institute of Science and Technology (KIST)

Abstract

Feed-forward 3D Gaussian Splatting (3DGS) reconstructs a scene in a single forward pass, replacing per-scene optimization with a network trained across many scenes. Its quality, however, degrades sharply as the number of input images drops. The bottleneck is upstream of the reconstruction heads: from a few unposed views, the internal representation they read carries no evidence for unobserved regions, leaving holes, floaters, and blur. The common remedy supplies that evidence as pixels, synthesizing extra views with an image or video generator and re-encoding them, which is costly and not 3D-consistent by construction. We instead densify the evidence itself. We present DensiTok, a plug-in module for pretrained feed-forward 3DGS models that densifies their internal geometry tokens directly, making a frozen backbone behave as though it had observed many more views than it was given. DensiTok compresses those tokens into a compact latent space, completes the latents of the unobserved viewpoints in a single flow-matching step conditioned on camera geometry, and decodes them back into tokens that the original reconstruction heads. The same module design can be integrated into different pretrained predictors while keeping each backbone and its reconstruction heads frozen. Completion in a low-dimensional latent space requires no image synthesis or additional encoder passes. Across three pretrained backbones and two benchmarks, DensiTok consistently improves sparse-view reconstruction and recovers much of the gap to dense-view reconstruction.

Figures & tables

Figure 1: DensiTok augments geometry tokens for unobserved views while keeping the pretrained encoder and reconstruction heads frozen. With only three input views, it enables AnySplat to exceed its 16-view baseline in PSNR on RealEstate10K.

Figure 2: Geometry token compression and completion. Stage 1 trains a VAE with Level-Adaptive Modulation (LAM-VAE) to compress and reconstruct geometry tokens. level-adaptive fusion and the VAE encoder together form the LAM-VAE encoder, with a level-wise residual into the posterior mean. Stage 2 freezes the VAE and trains a camera-conditioned DiT to complete latents for unobserved views. The geometry encoder remains frozen throughout, while the reconstruction heads remain frozen or undergo minimal appearance adaptation.

Dense-view ( $N=16$ )
Methods	RealEstate10K ( Zhou et al., 2018 )							DL3DV ( Ling et al., 2024 )							Runtime
	Appearance			Camera		Depth		Appearance			Camera		Depth		Runtime
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$	(s) $\downarrow$
AnySplat ( Jiang et al., 2025 )	24.04	0.792	0.189	20.4	77.5	0.835	0.151	22.30	0.683	0.246	24.1	85.4	0.803	0.158	0.434
WorldMirror ( Liu et al., 2025 )	25.11	0.856	0.161	41.8	83.3	0.880	0.110	24.47	0.819	0.173	69.0	94.0	0.825	0.141	0.495
Sparse-view ( $n=2$ )

Table 1: Quantitative results on RealEstate10K and DL3DV with

n\in\{2,3,5,7\}

sparse input views and dense-view references (

N=16

Figure 3: Qualitative novel-view RGB and depth comparisons for AnySplat with and without DensiTok on RE10K and DL3DV, using $n=2$ input views (left) and $n=3$ input views (right).

	RE10K							DL3DV
LAM-VAE encoder	PSNR	SSIM	LPIPS	AUC@ $3^{\circ}$	AUC@ $30^{\circ}$	$\delta_{1.25}$	AbsRel	PSNR	SSIM	LPIPS	AUC@ $3^{\circ}$	AUC@ $30^{\circ}$	$\delta_{1.25}$	AbsRel
Uniform fusion	19.76	0.673	0.266	18.3	76.3	0.808	0.184	19.53	0.611	0.314	21.7	85.2	0.760	0.199
MLP fusion	21.19	0.725	0.243	20.8	77.4	0.831	0.164	20.53	0.644	0.291	23.8	86.1	0.781	0.181
+ level-adaptive fusion	22.11	0.756	0.228	25.0	78.1	0.836	0.159	21.89	0.684	0.267	28.8	87.6	0.787	0.177
+ level-wise residual	21.86	0.747	0.234	24.2	78.6	0.841	0.156	21.56	0.675	0.274	27.5	88.2	0.794	0.173
+ both	22.56	0.766	0.223	26.7	78.9	0.846	0.153	22.40	0.699	0.255	30.4	88.9	0.801	0.170

Table 2: Ablations of Level-Adaptive Modulation with AnySplat and

n=2

observed views.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Architecture of a DiT block for geometry latent augmentation. Time and camera embeddings produce per-token AdaLN shift, scale, and gate parameters for the attention and MLP branches. Frame and global blocks differ only in their attention scope.

Setting	VAE encoder	VAE decoder	DiT
Hidden channels	512	512	768
Frame/global block pairs	6	6	12
Total transformer blocks	12	12	24
Attention heads	8	8	12

Appendix

Table 4: Shared VAE and DiT architecture for all three backbones. Each pair contains a frame block followed by a global block.

Setting	VAE	DiT	One-step finetuning
Optimizer	AdamW	AdamW	AdamW
Maximum learning rate	$2\times 10^{-4}$	$2\times 10^{-4}$	$4\times 10^{-5}$
Minimum learning rate	$10^{-6}$	$2\times 10^{-6}$	$2\times 10^{-6}$
Learning-rate schedule	Cosine decay	Cosine decay	Cosine decay
Optimizer betas	$(0.9,0.95)$	$(0.9,0.95)$	$(0.9,0.95)$
Weight decay	$0.01$	$0.01$	$0.01$

Appendix

Table 5: Training configurations for the three stages of the main AnySplat experiments.

Dense-view ( $N=16$ )
Methods	RealEstate10K ( Zhou et al., 2018 )							DL3DV ( Ling et al., 2024 )							Runtime
	Appearance			Camera		Depth		Appearance			Camera		Depth		Runtime
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$	(s) $\downarrow$
DA3 ( Lin et al., 2025 )	21.28	0.712	0.249	43.8	86.2	0.538	0.346	22.73	0.717	0.222	87.8	97.2	0.556	0.322	0.514
Sparse-view ( $n=2$ )
DA3	19.69	0.667	0.300	55.8	89.6	0.369	0.471	20.86	0.646	0.287	88.3	98.8	0.402	0.452	0.112

Appendix

Table 6: Quantitative results for DA3 on RealEstate10K and DL3DV with

n\in\{2,3,5,7\}

sparse input views and dense-view references (

N=16

). Camera AUC is reported in percent. Runtime is reported in seconds in a shared column.

Figure 5: Qualitative novel-view RGB and depth comparisons on RE10K and DL3DV with $n=2$ input views (left) and $n=3$ input views (right), each with and without DensiTok. Red boxes mark the enlarged RGB and depth crops. GT denotes the ground-truth target image.

Figure 7: Visualization of Gaussian reconstructions and camera poses for AnySplat with and without DensiTok. The top row shows the original AnySplat backbone and the bottom row shows DensiTok. Each scene is paired with a camera visualization, where red denotes ground truth and blue denotes predictions. DensiTok produces clearer scene structure and closer agreement with ground-truth camera positions and viewing orientations.

RE10K

DL3DV

Input cameras

Appearance

Depth

Appearance

Depth

Method

Intr.

Extr.

PSNR

SSIM

LPIPS

\delta_{1.25}

AbsRel

PSNR

SSIM

LPIPS

\delta_{1.25}

AbsRel

Runtime

\uparrow

\uparrow

\downarrow

\uparrow

\downarrow

\uparrow

\uparrow

\downarrow

\uparrow

\downarrow

(s)

\downarrow

Sparse-view (

n=2

)

pixelSplat

\circ

\circ

26.125

0.864

0.131

0.608

0.359

23.358

0.721

0.239

0.489

0.467

0.088

MVSplat

\circ

\circ

26.319

0.865

0.127

0.769

0.220

22.894

0.704

0.235

0.687

0.275

0.039

Appendix

Table 7: Comparison with sparse-view feed-forward Gaussian Splatting methods on RE10K and DL3DV at

256\times 256

with

n\in\{2,3,5,7\}

. NoPoSplat results are included for

n\in\{2,3\}

. Intr. and Extr. indicate whether the evaluated configuration requires externally supplied input camera intrinsics and extrinsics for reconstruction. Orange and yellow mark the best and second-best values for each metric at the same

n

, respectively. Rankings are determined before rounding, with ties sharing the same color.

Figure 8: Qualitative novel-view RGB comparisons with sparse-view feed-forward Gaussian Splatting models. DensiTok uses the AnySplat backbone and requires neither input camera intrinsics nor extrinsics. Methods are grouped by the input camera information they require. The leftmost column shows two input views for each scene, and GT denotes the ground-truth target image. Red boxes mark regions enlarged below each rendering.

Sparse-view ( $n=2$ )
	RE10K			DL3DV
Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	Runtime
	$\uparrow$	$\uparrow$	$\downarrow$	$\uparrow$	$\uparrow$	$\downarrow$	(s) $\downarrow$
NVComposer	14.897	0.528	0.508	12.078	0.359	0.610	48.502
Matrix3D	20.934	0.700	0.284	18.149	0.523	0.376	25.592
GLD	23.608	0.768	0.219	21.440	0.621	0.270	76.824

Appendix

Table 8: Comparison with diffusion-based novel-view synthesis methods on RE10K and DL3DV at

512\times 512

with

n\in\{2,3,5,7\}

. Runtime is reported in seconds in a shared column. Orange and yellow mark the best and second-best values for each metric at the same

n

, respectively. Rankings are determined before rounding, with ties sharing the same color.

Figure 9: Qualitative novel-view RGB comparisons of DensiTok with GLD, Matrix3D, and NVComposer using $n=2$ input views. The leftmost column shows the input views for each scene, and GT denotes the ground-truth target image. Red boxes mark regions enlarged below each rendering to compare the preservation of scene content and local geometry.

ETH3D ( Schops et al., 2017 )
$n$	Methods	Appearance			Camera		Depth
$n$	Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$
$2$	AnySplat	15.99	0.462	0.489	27.3	89.4	0.767	0.148
$2$	+ DensiTok	17.01	0.531	0.466	33.3	90.3	0.771	0.135
$3$	AnySplat	16.79	0.487	0.461	29.3	88.0	0.831	0.119
$3$	+ DensiTok	17.92	0.554	0.447	31.3	89.2	0.837	0.116

Appendix

Table 9: Generalization to ETH3D, ScanNet, and Mip-NeRF 360 with the AnySplat backbone and

n\in\{2,3,5,7\}

observed views. Camera AUC is reported in percent. Depth is evaluated against official ground truth on ETH3D and ScanNet and MoGe-2 predictions on Mip-NeRF 360.

		RE10K			DL3DV			Efficiency
Method	NFE	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	Runtime
	$\downarrow$	$\uparrow$	$\uparrow$	$\downarrow$	$\uparrow$	$\uparrow$	$\downarrow$	(s) $\downarrow$
Deterministic regression	1	20.86	0.712	0.246	19.97	0.631	0.288	0.258
One-step training from scratch	1	21.67	0.746	0.240	21.28	0.670	0.285	0.258
FM pretraining only	1	21.56	0.743	0.243	21.09	0.665	0.289	0.258
	8	21.84	0.755	0.236	21.59	0.676	0.279	0.990

Appendix

Table 10: Effect of one-step finetuning with AnySplat,

d_{z}=4

n=2

, and

N=16

. FM denotes flow matching and NFE denotes the number of DiT forward evaluations. Runtime is reported in a shared column.

		RE10K							DL3DV
		Appearance			Camera		Depth		Appearance			Camera		Depth
$n$	$d_{z}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$
$2$	$4$	22.56	0.766	0.223	26.7	78.9	0.846	0.153	22.40	0.699	0.255	30.4	88.9	0.801	0.170
$2$	$16$	22.64	0.770	0.215	27.5	79.0	0.854	0.151	22.24	0.700	0.247	39.6	90.8	0.804	0.164
$3$	$4$	24.28	0.813	0.184	25.8	79.7	0.853	0.146	24.28	0.755	0.205	34.4	89.6	0.809	0.158
$3$	$16$	24.57	0.822	0.171	27.1	79.5	0.862	0.145	24.29	0.760	0.194	38.5	90.5	0.812	0.155

Appendix

Table 11: Effect of latent channel count with AnySplat,

N=16

, and one-step inference.

		RE10K							DL3DV
		Appearance			Camera		Depth		Appearance			Camera		Depth
$n$	Gaussian head	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	AUC@ $3^{\circ}\uparrow$	AUC@ $30^{\circ}\uparrow$	$\delta_{1.25}\uparrow$	AbsRel $\downarrow$
$2$	Fully frozen	22.14	0.754	0.233	26.4	78.8	0.833	0.161	22.01	0.685	0.267	30.6	88.8	0.790	0.179
	Image embedding only	22.56	0.766	0.223	26.7	78.9	0.846	0.153	22.40	0.699	0.255	30.4	88.9	0.801	0.170
	Full finetuning	22.88	0.775	0.216	26.8	78.8	0.854	0.148	22.69	0.708	0.248	30.2	88.8	0.807	0.166
$3$	Fully frozen	24.07	0.806	0.190	25.6	78.8	0.846	0.150	24.05	0.747	0.212	34.1	88.8	0.803	0.163

Appendix

Table 12: Effect of Gaussian head finetuning with AnySplat,

d_{z}=4

N=16

, and one-step inference. Image embedding only denotes finetuning

h_{\mathrm{img}}

while keeping the remaining pretrained Gaussian head weights frozen.

Explore similar work

TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens
Apr 16, 2026Jiawei Ren, Michal Jan Tyszkiewicz, Jiahui Huang +1Feed-Forward 3D Gaussian Splatting 3D Gaussian
NVIDIA
ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion
Jul 22, 2026In Cho, Jeonghwan Cho, Mijin Yoo +2Feed-Forward 3D Gaussian Splatting Feed-Forward 3D Gaussian
Yonsei University, South Korea · National University of Singapore, Singapore
GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens
Apr 16, 2026Roni Itkin, Noam Issachar, Yehonatan Keypur +3Feed-Forward 3D Gaussian Splatting Feed-Forward 3D Gaussian
The Hebrew University of Jerusalem · Westlake University

Apr 16, 2026, cs.CVJ/K move · Enter open · S save

TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens

Jiawei Ren, Michal Jan Tyszkiewicz, Jiahui Huang, Zan Gojcic

NVIDIA

Feed-Forward 3D Gaussian Splatting 3D Gaussian Compact Latent Sensor Tokens Pixels Lotteries

Open paper AskPDF arXiv

In this work, we revisit several key design choices of modern Transformer-based approaches for feed-forward 3D Gaussian Splatting (3DGS) prediction. We argue that the common practice of regressing Gaussian means as depths along camera rays is suboptimal, and instead propose to directly regress 3D mean coordinates using only a self-supervised rendering loss. This formulation allows us to move from the standard encoder-only design to an encoder-decoder architecture with learnable Gaussian tokens, thereby unbinding the number of predicted primitives from input image resolution and number of views. Our resulting method, TokenGS, demonstrates improved robustness to pose noise and multiview inconsistencies, while naturally supporting efficient test-time optimization in token space without degrading learned priors. TokenGS achieves state-of-the-art feed-forward reconstruction performance on both static and dynamic scenes, producing more regularized geometry and more balanced 3DGS distribution, while seamlessly recovering emergent scene attributes such as static-dynamic decomposition and scene flow.

Abstract

Explore similar work

TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens

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

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