CATSplat: Context-Aware Transformer with Spatial Guidance for Generalizable 3D Gaussian Splatting from A Single-View Image

Wonseok Roh; Hwanhee Jung; Jong Wook Kim; Seunggwan Lee; Innfarn Yoo; Andreas Lugmayr; Seunggeun Chi; Karthik Ramani

cs.CVDec 17, 2024

CATSplat: Context-Aware Transformer with Spatial Guidance for Generalizable 3D Gaussian Splatting from A Single-View Image

Authors: Wonseok Roh, Hwanhee Jung, Jong Wook Kim, Seunggwan Lee, Innfarn Yoo, Andreas Lugmayr, Seunggeun Chi, Karthik Ramani, +1 more

Organizations: Korea University · Google · Purdue University

Abstract

Recently, generalizable feed-forward methods based on 3D Gaussian Splatting have gained significant attention for their potential to reconstruct 3D scenes using finite resources. These approaches create a 3D radiance field, parameterized by per-pixel 3D Gaussian primitives, from just a few images in a single forward pass. However, unlike multi-view methods that benefit from cross-view correspondences, 3D scene reconstruction with a single-view image remains an underexplored area. In this work, we introduce CATSplat, a novel generalizable transformer-based framework designed to break through the inherent constraints in monocular settings. First, we propose leveraging textual guidance from a visual-language model to complement insufficient information from a single image. By incorporating scene-specific contextual details from text embeddings through cross-attention, we pave the way for context-aware 3D scene reconstruction beyond relying solely on visual cues. Moreover, we advocate utilizing spatial guidance from 3D point features toward comprehensive geometric understanding under single-view settings. With 3D priors, image features can capture rich structural insights for predicting 3D Gaussians without multi-view techniques. Extensive experiments on large-scale datasets demonstrate the state-of-the-art performance of CATSplat in single-view 3D scene reconstruction with high-quality novel view synthesis.

Figures & tables

Figure 1 : Overview of the generalizable 3D scene reconstruction pipeline. The feed-forward network creates a 3D radiance field using 3D Gaussians, all within an end-to-end differentiable system.

Figure 2 : We introduce CATSplat , a C ontext- A ware T ransformer with S patial Guidance for Generalizable 3D Gaussian Splatting from a single image. (a) Our two main priors, and (b) Examples of text descriptions (from the VLM) representing an input image.

Figure 3 : Overview of CATSplat framework. CATSplat takes an image $\mathcal{I}$ and predicts 3D Gaussian primitives $\{(\bm{\mu}_{j},\bm{\alpha}_{j},\bm{\Sigma}_{j},\bm{c}_{j})\}^{J}_{j}$ to construct a scene-representative 3D radiance field in a single forward pass. Our primary goal is to go beyond the finite knowledge inherent in single-view image features leveraging our two innovative priors. Through cross-attention layers, we enhance image features $F^{\mathcal{I}}_{i}$ to be highly informative by incorporating valuable insights: contextual cues from text features $F^{C}_{i}$ , and spatial cues from 3D point features $F^{S}_{i}$ .

Figure 4 : Detailed transformer pipeline. In the $i$ -th layer, we first operate cross-attention between $F_{i}^{\mathcal{I}}$ and $F_{i}^{C}$ , then proceed cross-attention with $F_{i}^{S}$ . We also use a ratio $\gamma$ to preserve visual information from $F_{i}^{\mathcal{I}}$ while incorporating extra cues from $F_{i}^{C}$ and $F_{i}^{S}$ .

	$n=5$ (frames)			$n=10$ (frames)			$n=\textit{Random}$ (frames)
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
MPI [ 54 ]	27.10	0.870	–	24.40	0.812	–	23.52	0.785	–
BTS [ 60 ]	–	–	–	–	–	–	24.00	0.755	0.194
Splatter Image [ 51 ]	28.15	0.894	0.110	25.34	0.842	0.144	24.15	0.810	0.177
MINE [ 26 ]	28.45	0.897	0.111	25.89	0.850	0.150	24.75	0.820	0.179
Flash3D [ 50 ]	28.46	0.899	0.100	25.94	0.857	0.133	24.93	0.833	0.160

Table 1: Comparisons of Novel View Synthesis (NVS) performance with state-of-the-art single-view 3D reconstruction approaches on the RealEstate10K [ 72 ] dataset. Following the standard protocol from [ 26 , 50 ] , we evaluate NVS metrics on unseen target frames located

n

frames away from the input source frame. Also, we randomly sample an extra target frame within 30 frames apart from the source frame.

			RE10K Interpolation			RE10K Extrapolation
Input	Method	Framework	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Two-View	pixelNeRF [ 65 ]	NeRF	20.51	0.592	0.550	20.05	0.575	0.567
	Du et al . [ 14 ]	NeRF	24.78	0.820	0.213	21.83	0.790	0.242
	pixelSplat [ 8 ]	3DGS	26.09	0.864	0.136	21.84	0.777	0.216
	latentSplat [ 58 ]	3DGS	23.93	0.812	0.164	22.62	0.777	0.196
	MVSplat [ 10 ]	3DGS	26.39	0.869	0.128	23.04	0.813	0.185

Table 2: Comparisons of NVS performance with state-of-the-art few-view 3D reconstruction approaches on the RealEstate10K [ 72 ] . Although we mainly focus on comparing with the leading single-view method, Flash3D [ 50 ] , we also provide scores of two-view methods for additional references. Following Flash3D, we use interpolation and extrapolation protocols from previous works, [ 8 ] and [ 58 ] , respectively.

Cross Dataset	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
RE10K $\rightarrow$ NYUv2	Flash3D [ 50 ]	25.09	0.775	0.182
RE10K $\rightarrow$ NYUv2	CATSplat (Ours)	25.57	0.781	0.157
RE10K $\rightarrow$ ACID	Flash3D [ 50 ]	24.28	0.730	0.263
RE10K $\rightarrow$ ACID	CATSplat (Ours)	24.73	0.739	0.250
RE10K $\rightarrow$ KITTI	Flash3D [ 50 ]	21.96	0.826	0.132
RE10K $\rightarrow$ KITTI	CATSplat (Ours)	22.43	0.833	0.122

Table 3 : Comparisons of cross-dataset generalization with the state-of-the-art single-view 3DGS method, Flash3D [ 50 ] , on various real-world datasets: NYUv2 [ 47 ] , ACID [ 29 ] , and KITTI [ 17 ] .

Figure 5 : Ablation study to see the effect of iteratively incorporating our novel priors on the RE10K [ 72 ] ( $n$ = Random ). For clear ablations, we keep the number of entire transformer layers consistent across the experiments and adjust only the number of cross-attentions (CA).

	$n=\text{10}$ (frames)			$n=\textit{Random}$ (frames)
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Baseline	26.04	0.857	0.132	25.02	0.834	0.159
w/ Contextual	26.40	0.864	0.127	25.40	0.838	0.153
w/ Spatial	26.38	0.864	0.127	25.42	0.837	0.153
CATSplat	26.44	0.866	0.125	25.45	0.841	0.151

Table 4 : Ablation study to explore the effect of our two intelligent priors (Contextual and Spatial) across three different settings, as in Tab. 1 , on the RE10K [ 72 ] dataset. Here, the “Baseline” indicates our basic transformer architecture without any proposed priors.

Figure 6 : Qualitative comparisons of NVS performance between Flash3D [ 50 ] and ours with Ground Truth on the novel view frames from RealEstate10K [ 72 ] and ACID [ 29 ] (cross-dataset). We provide more visual results and details of user study in the supplementary material.

	$n=\text{10}$ (frames)			$n=\textit{Random}$ (frames)
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
w/ ConvNeXt-B	26.15	0.856	0.132	25.09	0.832	0.158
w/ ConvNeXt-L	26.17	0.857	0.132	25.12	0.833	0.157
w/ DINOv2-B	26.17	0.858	0.131	25.11	0.833	0.157
w/ DINOv2-g	26.19	0.859	0.131	25.17	0.834	0.156
w/ Contextual	26.40	0.864	0.127	25.40	0.838	0.153

Table 5 : Ablation study to see the effect of our novel priors versus visual priors from large-scale pre-trained image encoders, such as DINOv2 [ 39 ] and ConvNeXt V2 [ 61 ] , on the RE10K [ 72 ] dataset.

	$n=\text{10}$ (frames)			$n=\textit{Random}$ (frames)
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Baseline	26.04	0.857	0.132	25.02	0.834	0.159
w/ Scene Type	26.14	0.859	0.130	25.13	0.835	0.158
w/ Object List	26.23	0.862	0.128	25.25	0.836	0.155
w/ Extended	26.31	0.862	0.128	25.29	0.837	0.154
w/ Single Sent.	26.40	0.864	0.127	25.40	0.838	0.153

Table 6 : Ablation study to see the impact of different text description formats on generalizable tasks. The “Baseline” is as in Tab. 4 .

	$n=\text{10}$ (frames)			$n=\textit{Random}$ (frames)
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Baseline	26.04	0.857	0.132	25.02	0.834	0.159
w/o Depth Conc.	25.91	0.855	0.134	24.82	0.827	0.165
w/ Point Conc.	26.06	0.857	0.132	25.04	0.834	0.158
w/ Depth Feat.	26.18	0.859	0.130	25.16	0.835	0.157
w/ Point Feat.	26.38	0.864	0.127	25.42	0.837	0.153

Table 7 : Ablation study to explore strategies for enriching geometric knowledge from a single image. The “Baseline” is as in Tab. 4 .

	RE10K [ 72 ]		ACID [ 29 ]
Method	Preference ( $\%$ )	Likert $\uparrow$	Preference ( $\%$ )	Likert $\uparrow$
Flash3D [ 50 ]	11.58 $\pm$ 1.09	4.56 $\pm$ 0.30	8.59 $\pm$ 0.63	4.14 $\pm$ 0.21
CATSplat (Ours)	88.42 $\pm$ 1.09	6.04 $\pm$ 0.22	91.41 $\pm$ 0.63	5.27 $\pm$ 0.18

Table 8 : User study comparisons. We report mean preference percentage and a 7-point Likert scale with a 95% confidence interval.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1 : Examples of two types of user study questions. The first type of question (above) asks about preference between ours and Flash3D [ 50 ] , and the second (below) requires participants to rate the visual quality of the rendered image compared to the target.

Figure 2 : Detailed architecture of 3D point feature extraction from a monocular input image $\mathcal{I}$ . Our point cloud encoder takes back-projected points $P$ and produces point features $F^{S}$ based on the PointNet [ 42 ] structure. Here, T-Net indicates an affine transform network.

Figure 3 : Examples of input images with their corresponding estimated depth maps and back-projected 3D point clouds. For better visualization, we also present 3D point clouds with RGB colors.

Method	Training (hrs)	Inference (secs)	$n=\textit{Random}$ (frames)
Method	Training (hrs)	Inference (secs)	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Flash3D [ 50 ]	23.2	0.327	24.93	0.833	0.160
CATSplat (Ours)	29.1	0.393	25.45	0.841	0.151

Appendix

Table 1 : Comparison of computational costs with the state-of-the-art single-view 3DGS method, Flash3D [ 50 ] , on the RE10K [ 72 ] .

Method			$n=\textit{Random}$ (frames)
Baseline	Contextual	Spatial	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$\checkmark$	-	-	25.11	0.775	0.178
$\checkmark$	$\checkmark$	-	25.51	0.779	0.163
$\checkmark$	-	$\checkmark$	25.48	0.778	0.165
$\checkmark$	$\checkmark$	$\checkmark$	25.57	0.781	0.157

Appendix

Table 2 : Ablation study to investigate the effect of our two intelligent priors on the NYUv2 [ 47 ] dataset in cross-dataset settings.

Method			$n=\textit{Random}$ (frames)
Baseline	Contextual	Spatial	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$\checkmark$	-	-	24.26	0.732	0.261
$\checkmark$	$\checkmark$	-	24.57	0.735	0.253
$\checkmark$	-	$\checkmark$	24.62	0.737	0.254
$\checkmark$	$\checkmark$	$\checkmark$	24.73	0.739	0.250

Appendix

Table 3 : Ablation study to investigate the effect of our two intelligent priors on the ACID [ 72 ] dataset in cross-dataset settings.

Figure 4 : Examples of four different formats of text descriptions from the VLM [ 31 ] , as described in Tab.5 in the main paper.

	$n=\text{10}$ (frames)			$n=\textit{Random}$ (frames)
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
OpenFlamingo	26.08	0.858	0.131	25.06	0.832	0.158
BLIP2 T5	26.29	0.860	0.129	25.27	0.833	0.156
LLaVA 7B	26.19	0.861	0.129	25.23	0.834	0.156
LLaVA 13B	26.40	0.864	0.127	25.40	0.838	0.153

Appendix

Table 4 : Ablation study to see the impact of text embeddings from different VLMs: OpenFlamingo [ 5 ] , BLIP2 [ 27 ] , and LLaVA [ 31 ] .

Figure 6 : Qualitative comparisons with various novel view synthesis approaches, including MASt3R [ 25 ] , Splatt3R [ 49 ] , MoGe [ 56 ] , and ZeroNVS [ 45 ] . All methods are fairly evaluated on unseen scenes in cross-dataset settings across ACID [ 29 ] and NYUv2 [ 47 ] datasets.

Figure 7 : Qualitative comparisons of 3D reconstruction between Flash3D [ 50 ] and ours with Ground Truth. We visualize zoom-in views of 3D Gaussian distributions and depth maps from them.

Figure 8 : Failure cases of CATSplat. When invisible areas in the input become visible in the target, ours might be less productive.

Figure 9 : Qualitative comparisons between Flash3D [ 50 ] and Ours with Input Image and Ground Truth on the RealEstate10K [ 72 ] dataset.

Figure 10 : Qualitative comparisons between Flash3D [ 50 ] and Ours with Input Image and Ground Truth on the RealEstate10K [ 72 ] dataset.

Figure 11 : Qualitative comparisons between Flash3D [ 50 ] and Ours with Input Image and Ground Truth on the ACID [ 29 ] dataset.

Figure 12 : Qualitative comparisons between Flash3D [ 50 ] and Ours with Input Image and Ground Truth on the KITTI [ 17 ] dataset.

Explore similar work

Aug 7, 2026cs.CV

InstanceSplat: Instance-Aware Feed-Forward 3D Gaussian Splatting for Scene Understanding

Feed-forward 3D Gaussian Splatting (3DGS) enables efficient and generalizable 3D reconstruction, but current feed-forward 3DGS methods for scene understanding remain largely category-oriented. In contrast, instance-aware 3DGS methods typically rely on per-scene optimization and often decouple reconstruction from instance and semantic learning, limiting reciprocal interactions among them. We present InstanceSplat, a unified feed-forward 3DGS framework for generalizable 3D reconstruction and instance-aware scene understanding from pose-free multi-view images. In a single forward pass, InstanceSplat constructs an instance-aware Gaussian representation that jointly encodes appearance, geometry, instance identity, and language-aligned semantics. Shared 3D Gaussians ground instance identities across views, producing renderable and cross-view-consistent instance features. To allow reconstruction and scene understanding to benefit from each other, we further design an instance-centric learning strategy that connects reconstruction, instance learning, and semantic learning through shared instance structure. Specifically, instance cues guide reconstruction, language-aligned semantics strengthen the discrimination of confusing same-category instances, and instance regions aggregate semantic evidence into coherent object-level predictions. Experiments on novel-view synthesis, instance segmentation, and open-vocabulary semantic understanding under varying input-view settings and on an unseen dataset demonstrate state-of-the-art performance, practical efficiency, and strong generalization.

Minchao Jiang, Xiaoxuan Ma, Shunyu Jia +3

1Shanghai Jiao Tong University · 2Eastern Institute of Technology, Ningbo · 3Carnegie Mellon University +2

Feed-Forward 3D Gaussian Splatting 3D Scene Understanding 3D Reconstruction Scene Reasoning 3D Gaussian Primitives

Ask PDF

Apr 30, 2026cs.CV

Generalizable Sparse-View 3D Reconstruction from Unconstrained Images

Reconstructing 3D scenes from sparse, unposed images remains challenging under real-world conditions with varying illumination and transient occlusions. Existing methods rely on scene-specific optimization using appearance embeddings or dynamic masks, which requires extensive per-scene training and fails under sparse views. Moreover, evaluations on limited scenes raise questions about generalization. We present GenWildSplat, a feed-forward framework for sparse-view outdoor reconstruction that requires no per-scene optimization. Given unposed internet images, GenWildSplat predicts depth, camera parameters, and 3D Gaussians in a canonical space using learned geometric priors. An appearance adapter modulates appearance for target lighting conditions, while semantic segmentation handles transient objects. Through curriculum learning on synthetic and real data, GenWildSplat generalizes across diverse illumination and occlusion patterns. Evaluations on PhotoTourism and MegaScenes benchmark demonstrate state-of-the-art feed-forward rendering quality, achieving real-time inference without test-time optimization

Vinayak Gupta, Chih-Hao Lin, Shenlong Wang +2

University of Maryland, College Park · University of Illinois Urbana-Champaign · Johns Hopkins University

3D Reconstruction Sparse-View 3D Generation Visual Embeddings

Ask PDF

May 7, 2026cs.CV

Sparse-to-Complete: From Sparse Image Captures to Complete 3D Scenes

We introduce S2C-3D, a novel sparse-view 3D reconstruction framework for high-fidelity and complete scene reconstruction from as few as six to eight images. Our framework features three components: a specialized diffusion model for scene-specific image restoration, a training-free view-consistency conditioned sampling process in the diffusion model for refined Gaussian optimization, and a camera trajectory planning scheme to ensure comprehensive scene coverage. The specialized diffusion model is developed by finetuning a pretrained architecture on the input views and their corresponding degraded counterparts. The adaptation to the scene distribution allows the model to repair Gaussian renderings while effectively eliminating domain gaps. Meanwhile, the trajectory planning scheme optimizes scene coverage by connecting each newly sampled camera to its two nearest neighbors. By iteratively constructing paths and retaining only those that significantly enhance visibility, the scheme establishes a trajectory that covers the entire scene. To address multi-view conflicts, the view-consistency conditioned sampling process quantifies the consistency between neighboring repaired images. This information is injected as a condition into the sampling process of the frozen diffusion model, facilitating the generation of view-consistent images without additional training. Consequently, our approach produces high-fidelity 3D Gaussians that are robust to artifacts. Experimental results demonstrate that S2C-3D outperforms state-of-the-art methods, constructing high-quality scenes that are free from missing regions, blurring, or other artifacts with very sparse inputs. The source code and data are available at https://gapszju.github.io/S2C-3D.

Yiyang Shen, Yin Yang, Kun Zhou +1

State Key Lab of CAD&CG, Zhejiang University, China · University of Utah, USA · State Key Lab of CAD&CG, Zhejiang University, China and Hangzhou Research Institute of Holographic and AI Technology, China +1

3D Reconstruction 3D Generation Image Restoration Scene Understanding Camera Trajectories

Ask PDF