betaFeedback

WorksComputer Science

OctreeSplat: Hierarchical Octree-Guided Feed-Forward Gaussian Splatting

Abstract

Recent feed-forward 3D Gaussian Splatting (3DGS) methods typically regress pixel-aligned Gaussian primitives, often causing excessive overlap and artifacts, while inaccuracies in predicted camera poses can lead to misalignment in novel-view synthesis (NVS). We present VoxelTTO, a feed-forward framework for reconstructing geometrically accurate 3DGS from an arbitrary number of images and optional camera parameters. VoxelTTO aggregates dense image features into global voxels and decodes Gaussians from voxel features, breaking the pixel-to-Gaussian correspondence. To exploit known camera parameters while keeping the pretrained visual foundation model (VFM) parameters frozen, we introduce test-time optimization (TTO) that adapts lightweight LoRA modules using pose supervision. We further replace vanilla 3DGS rasterization with stochastic solid volume rendering during training and inference, improving geometric fidelity. Training updates only the voxel-aligned Gaussian reconstruction modules, requiring 80 GPU hours. Experiments on Replica, Tanks and Temples, RE10K, and DTU demonstrate improved RGB-D NVS and camera-pose alignment relative to prior methods.

VoxelTTO directly regresses a geometrically accurate Gaussian-splatting scene from posed images

Figure 1. VoxelTTO directly regresses a geometrically accurate Gaussian-splatting scene from posed images. It precisely aligns camera poses and improves RGB and depth rendering in both input and test views across indoor–outdoor and synthetic–realworld scenes.

Introduction

Reconstructing three-dimensional scenes from images is a fundamental computer-vision problem with applications in autonomous driving [52], virtual reality [2], and robotics [55]. Conventional 3D reconstruction systems from images rely on multistage pipelines comprising feature matching [9, 36] and structure from motion (SfM) [37]. Recent VFMs [44, 47] offer an end-to-end alternative, using a large Transformer to directly predict 3D attributes from arbitrary uncalibrated image collections.

Despite their strong geometric priors, point-map and depth-map representations produced by VFMs often fail to capture fine details and photorealistic appearance. By contrast, 3DGS [22], and its extensions [17, 29] use richer scene representations to deliver accurate NVS. However, these methods commonly recover camera parameters with SfM before performing costly per-scene optimization. Some approaches directly predict 3DGS scenes [5, 6, 53], but they typically require accurate camera parameters and can degrade in complex scenes. More recent methods [19, 27, 56] extend the feed-forward VFMs to 3DGS. They regress pixel-aligned 3DGS scenes directly from unposed images, but their geometric accuracy remains limited. Several VFMs encode camera parameters within the backbone to exploit additional priors and improve prediction accuracy [19, 21, 27]. Nevertheless, the resulting accuracy remains insufficient, and substantial misalignment persists during test-view synthesis.

In this work, we present VoxelTTO, a voxel-aligned feed-forward 3DGS framework with test-time optimization (TTO). VoxelTTO reconstructs geometrically accurate 3DGS scenes from arbitrary numbers of images and optional camera parameters. When camera priors are available, it supports accurate NVS from test views. To reduce excessive overlap and artifacts, VoxelTTO departs from conventional pixel-aligned feed-forward 3DGS, up-samples patch tokens from a frozen VFM backbone with DPT and fuses them into global voxels. A sparse convolutional U-Net encodes the voxel features, and an appearance-geometry decoupled self-splitting MLP regresses Gaussian primitive parameters. We replace 3DGS rasterization with stochastic solid volume rendering, which provides stronger geometric supervision during optimization and improves geometric fidelity for NVS .

To exploit camera priors, TTO inserts LoRA modules into the VFM backbone. Pose-supervised adaptation of these modules improves alignment between predicted and ground-truth camera parameters. During training, we freeze the VFM backbone and optimize only the Voxel-Aligned Feed-Forward 3DGS module. This design introduces 75.3M trainable parameters, significantly reduces training cost to 80 GPU hours, and is compatible with any VGGT-style VFM.

Our main contributions are as follows:

  • We propose VoxelTTO, a voxel-aligned feed-forward 3DGS reconstruction framework that recovers geometrically accurate scenes from arbitrary image collections and optional camera parameters.

  • We introduce voxel-aligned feed-forward 3DGS and stochastic solid volume rendering supervision to reduce excessive overlap and artifacts while strengthening geometry.

  • We introduce pose-supervised TTO to exploit camera priors and improve alignment.

  • Extensive experiments across multiple datasets demonstrate improved camera-pose estimation and more accurate RGB-D rendering in both input and test views compared to existing feed-forward 3DGS methods.

Related Works

Optimization-Based Novel-View Synthesis

Neural radiance fields (NeRF) [30] represent continuous scene density and view-dependent color with neural fields, enabling camera-level NVS through volume rendering. Subsequent work has improved reconstruction speed [32], anti-aliasing [3], and geometric quality [46, 49]. As volume rendering evaluates samples along each ray, NeRF-based methods typically require a significant amount of time. In contrast, 3DGS [22] represents a scene as anisotropic Gaussian primitives and achieves NVS at over 30 FPS. Several studies improve Gaussian representations [8, 17] or optimization strategies [13, 18] to recover geometry accurately. Recent work has established an equivalence between stochastic solid volume rendering [31] and 3DGS rasterization [12, 58], yielding strong depth-rendering and surface-reconstruction accuracy.

These optimization-based methods require accurate camera parameters, which are typically estimated using SfM pipelines[37, 38, 43, 47]. This requirement introduces preprocessing and can accumulate error. Several methods jointly optimize camera parameters and 3DGS scenes, but generally require sequential images [11, 20] or depth maps [14]. Vanilla 3DGS initializes primitives by KNN distances from a colored point cloud and optimizes each scene independently [22]. Subsequent methods directly regress a 3DGS scene from images and camera parameters without per-scene optimization [5, 6, 53].

Visual Foundation Models

Image-based reconstruction has traditionally relied on separate SfM and dense-reconstruction stages. Recent VFMs directly regress 3D scene information from multiple views. Dust3R [47] uses self and cross-attention to infer aligned point maps and camera parameters from image pairs. Subsequent work improves reconstruction accuracy [25], extends the formulation to multiple views [42], and accelerates inference [54]. VGGT alternates single image attention with global attention to regress depth maps, point maps, and camera parameters, from arbitrary views. Later work increases model accuracy[27, 45] and improves order invariance [50]. Several VFMs [19, 27, 56] further support direct regression of 3DGS. Their Gaussian primitives are usually pixel-aligned with depth maps, which inevitably introduces excessive overlap and rendering artifacts. Directly regressing Gaussians from fixed tokens [26, 35] alleviates this issue but commonly sacrifices geometric accuracy. Recent works begin to explore aligning Gaussian primitives with voxel-based representations[48].

VFM geometry is predicted across views and may not strictly satisfy multiview constraints. Test-time training (TTT) addresses the quadratic cost of query–key interactions in self-attention, enabling linear-complexity visual sequence modeling [15]. This strategy has also been extended to VFMs to replace quadratic global attention and improve scalability to large image collections [10]. Recent TCO inserts LoRA modules into attention layers and constrains them using known camera parameters and depth maps, improving pose estimation and dense reconstruction [60]. Unlike approaches that encode depth maps or camera parameters into the backbone [19, 21, 27], TCO adapts lightweight modules without changing the base VFM architecture. Moreover, the inferred camera parameters gmg_m can still differ substantially from ggtg_{\mathrm{gt}}. Even after Umeyama SIM(3) alignment [41], this error causes substantial camera misalignment in subsequent NVS.

Method

We propose VoxelTTO, a feed-forward method that predicts a 3DGS scene from multiview images and optional camera parameters. By aggregating dense features into voxels and processing them with a sparse-convolutional U-Net, VoxelTTO delivers high geometric fidelity and appearance performance. TTO further addresses the misalignment commonly observed when VGGT-style VFMs predict Gaussian-splatting scenes, improving accuracy in test views.

Preliminaries and Challenges

VFMs [44, 47] use a large Transformer Φη\Phi_{\eta} to learn cross-view visual representations from MM uncalibrated images I={Im}m=1M\mathcal{I} = \{\mathcal{I}_m\}_{m=1}^{M}. Separate prediction heads Γψ\Gamma_{\psi} recover 3D scene information, including point maps Pm∈RH×W×3\mathbf{P}_m \in\mathbb{R}^{H \times W \times3}, depth maps Dm∈RH×W\mathbf{D}_m \in\mathbb{R}^{H \times W}, and camera parameters gm\mathbf{g}_m, as shown in Equation (1).

(Pm,Dm,gm)=Γψ(Φη(I)),m=1,…,M(1)(\mathbf{P}_m,\mathbf{D}_m,\mathbf{g}_m) = \Gamma_{\psi}\left(\Phi_{\eta}(\mathcal{I})\right), \qquad m = 1,\ldots,M \tag*{(1)}

3DGS VFMs typically upsample patch tokens with DPT to obtain pixel-aligned Gaussian attributes (center μq\boldsymbol{\mu}_q, opacity σq\sigma_q, rotation quaternion rq\mathbf{r}_q, scale sq\mathbf{s}_q, and SH coefficients hq\mathbf{h}_q). Such methods often lack multiview consistency and create excessive overlap, producing artifacts. Direct token-to-Gaussian methods commonly have lower geometric accuracy [26].

To exploit camera priors, several VFMs [19, 21, 27] encode ground-truth camera parameters ggt\mathbf{g}_{\mathrm{gt}} as tokens before the backbone and incorporate them during inference. Although fusing auxiliary camera parameters with RGB features improves performance, it changes the architecture and prevents direct reuse of unconstrained VFMs. Moreover, the inferred camera parameters gm\mathbf{g}_m can still differ substantially from ggt\mathbf{g}_{\mathrm{gt}}. Even after Umeyama SIM(3) alignment [41], this error causes substantial camera misalignment in subsequent NVS.

Voxel-Aligned Feed-Forward 3DGS

Network Architecture

VoxelTTO uses DINOv2 [33] to extract multiview patch tokens and employs alternating attention to model the scene. We add LoRA to the alternating-attention layers, as detailed in subsection 3.3. The frozen VFM backbone first predicts a depth map Dm\mathbf{D}_m, camera parameters gm\mathbf{g}_m, and confidence map Km\mathbf{K}_m for each view. We add a feature DPT that upsamples patch tokens into dense features Fm∈RH×W×C\mathbf{F}_m \in\mathbb{R}^{H \times W \times C}. We then back-project pixels into world coordinates, obtaining a dense feature point cloud P={(pl,al,fl,κl)}l=1Npts\mathcal{P} = \{(\mathbf{p}_l,\mathbf{a}_l,\mathbf{f}_l,\kappa_l)\}_{l=1}^{N_{\mathrm{pts}}}, where NptsN_{\mathrm{pts}} is the number of points and pl\mathbf{p}_l, al\mathbf{a}_l, fl\mathbf{f}_l, and κl\kappa_l denote point coordinate, color, feature, and confidence. To reduce memory use and avoid excessive overlap, we aggregate the point cloud into voxel features. Let Pv\mathcal{P}_v denote the points within voxel vv. Its feature center is xv=∑l∈Pvκlpl/∑l∈Pvκl\mathbf{x}_v = \sum_{l \in\mathcal{P}_v} \kappa_l \mathbf{p}_l / \sum_{l \in\mathcal{P}_v} \kappa_l, which is not constrained to the regular voxel center and thus preserves geometric accuracy. Voxel color av\mathbf{a}_v and feature fv\mathbf{f}_v are also confidence-weighted averages. We define voxel confidence as κv=min⁡l∈Pvκl\kappa_v = \min_{l \in\mathcal{P}_v} \kappa_l to reduce the impact of unreliable points. A 3D U-Net with sparse convolutions [7] refines voxel features for scene understanding while avoiding computation in empty voxels. Additional implementation details are provided in the supplementary material.

To increase the representational capacity of each voxel, we decode every voxel feature fv\mathbf{f}_v into nn Gaussian primitives, allowing one voxel to represent complex local detail. To avoid entangling geometry and appearance during optimization [26], we use separate geometry and appearance MLP branches, Dgeo\mathcal{D}_{\mathrm{geo}} and Dapp\mathcal{D}_{\mathrm{app}}. Centers, rotation quaternions, scales, and opacities of nn Gaussian primitives are decoded by Dgeo\mathcal{D}_{\mathrm{geo}}. Second-order spherical-harmonic coefficients are decoded by Dapp\mathcal{D}_{\mathrm{app}}.

Stochastic Solid Volume Rendering

To further improve the geometric accuracy of NVS, we use stochastic solid volume rendering[31] instead of vanilla 3DGS rasterization. With an appropriate vacancy function vv, volume rendering of the stochastic solid is equivalent to 3DGS rasterization [58], as shown in Equation (2), where ω\boldsymbol{\omega} is the ray direction, σ(x,ω)\sigma(\mathbf{x},\boldsymbol{\omega}) is the stochastic solid attenuation coefficient at point x\mathbf{x} along the ray, T(t)T(t) is accumulated transmittance at position tt, and c\mathbf{c} is the directional color.

Overview of the VoxelTTO pipeline

Figure 2. Overview of the VoxelTTO pipeline.

σ(x,ω)=∣ω⊤∇log⁡v(x)∣=∣ω⊤∇v(x)∣v(x)\sigma(\mathbf{x},\boldsymbol{\omega})=\left|\boldsymbol{\omega}^{\top}\nabla\log v(\mathbf{x})\right|=\frac{\left|\boldsymbol{\omega}^{\top}\nabla v(\mathbf{x})\right|}{v(\mathbf{x})}
T(t)=exp⁡(−∫tntσ(x(s),ω) ds)(2)T(t)=\exp\left(-\int_{t_n}^{t}\sigma(\mathbf{x}(s),\boldsymbol{\omega})\,ds\right) \tag*{(2)}
C=∫tntfT(t)σ(x(t),ω)c(x(t),ω) dt\mathbf{C}=\int_{t_n}^{t_f}T(t)\sigma(\mathbf{x}(t),\boldsymbol{\omega})\mathbf{c}(\mathbf{x}(t),\boldsymbol{\omega})\,dt

An appropriate vacancy vv is required to obtain images equivalent to 3DGS rasterization. GGGS[58] proves that the renderings are equivalent when v(x)=1−G(x)v(\mathbf{x})=\sqrt{1-G(\mathbf{x})}. The accumulated transmittance is then computed as shown in Equation (3), where ti∗t_i^* is the location of the Gaussian maximum along the ray.

Ti(t)={1−Gi(t),t≤ti∗,1−Gi(ti∗)1−Gi(t),t>ti∗.(3)T_i(t)= \begin{cases} \sqrt{1-G_i(t)}, & t\leq t_i^*,\\ \dfrac{1-G_i(t_i^*)}{\sqrt{1-G_i(t)}}, & t>t_i^*. \end{cases} \tag*{(3)}

For depth rendering, we use the median depth, defined by the point where T(tmed)=0.5T(t_{\mathrm{med}})=0.5 with a fixed number of bisection iterations. During backpropagation, we differentiate the fixed condition T(tmed)=0.5T(t_{\mathrm{med}})=0.5 with respect to Gaussian parameters θ\theta, as shown in Equation (4).

∂T∂t∂tmed∂θ+∂T∂θ=0(4)\frac{\partial T}{\partial t}\frac{\partial t_{\mathrm{med}}}{\partial\theta}+\frac{\partial T}{\partial\theta}=0 \tag*{(4)}

As shown in Figure 3, stochastic solid volume rendering produces RGB images that are nearly identical to 3DGS rasterization for the same Gaussian scene. For depth rendering, however, it offers a clear geometric advantage by reducing holes and blurred boundaries.

Geometric advantages of stochastic solid volume rendering over 3DGS rasterization

Figure 3. Geometric advantages of stochastic solid volume rendering over 3DGS rasterization.

Test-Time Optimization for VFMs

For reconstruction from posed images, some VFMs encode camera parameters with an MLP and inject them directly into the Transformer. Although this approach improves geometric accuracy, the improvement is insufficient for test-view NVS. As shown in Figure 4, AnySplat maintains reasonable input view quality despite artifacts, but produces misalignment in test views.

Test-view misalignment after injecting cameras

Figure 4. Test-view misalignment after injecting cameras.

To address this issue, we introduce LoRA [16] into the VFM backbone and perform TTO. Following the DA3 architecture [27], our network has 40 attention layers. Layers 1–12 use single-view attention, whereas layers 13–40 alternate between single-view and global attention. We add LoRA to the Q, K, and V projections in layers 13–40. Each layer shares a down-projection matrix A, while their up-projection matrices B are independent.

When aligned camera parameters are provided, we compute the optional TTO loss after camera decoding and update the LoRA parameters by backpropagation. After fixed iterations, the backbone produces more accurate camera parameters aligned with ground-truth cameras. VGGT-style VFMs usually use the first image as the scene coordinate frame, and image-based reconstruction has scale ambiguity. We therefore align poses before calculating the loss. We fix the first ground-truth camera pose to the identity matrix. The rotation loss is constructed from the relative rotation between predicted rotations RipR_i^p and ground-truth rotations RigR_i^g, as shown in Equation (5).

Lrot=12M∑i=1M[3−tr⁡((Rip)TRig)].(5)\mathcal{L}_{\mathrm{rot}} = \frac{1}{2M}\sum_{i=1}^{M}\left[3-\operatorname{tr}\left((R_i^p)^{\mathsf{T}}R_i^g\right)\right]. \tag*{(5)}

To address scale ambiguity, we normalize each translation set by mean scale s=M−1∑i=1M∥ti∥2s=M^{-1}\sum_{i=1}^{M}\lVert t_i\rVert_2, computed separately for the predicted and ground-truth translations. The scale-normalized translation loss is then calculated by SmoothL1. Camera intrinsics are also supervised. The VFM parameterizes intrinsics as k=(fx,fy,cx,cy)\mathbf{k}=(f_x,f_y,c_x,c_y), and we directly apply SmoothL1 to the predicted and ground-truth values. The camera-based TTO objective is LTTO=λpose(Lrot+wtLtrans)+λKLintrinsics\mathcal{L}_{\mathrm{TTO}}=\lambda_{\mathrm{pose}}\left(\mathcal{L}_{\mathrm{rot}}+w_t\mathcal{L}_{\mathrm{trans}}\right)+\lambda_K\mathcal{L}_{\mathrm{intrinsics}}.

Training

Datasets

We train VoxelTTO on ScanNet++ [57], Infinigen [34], ARKitScenes [4], and DL3DV-10K [28]. These datasets span indoor and outdoor environments as well as synthetic and real-world scenes. For each dataset, we define a minimum and maximum frame gap, randomly select the interval between the first and last frame within that range, and sample frames from it.

Training Details

Our architecture follows DA3, uses L=40L=40 alternating-attention layers as shown in subsection 3.3, initialized with pretrained weights [27]. The full model has 1.39B parameters, of which only 75.3M parameters are trainable. We optimize for 30K iterations with Adam, an initial lr of 3×10−63\times10^{-6}, cosine annealing, 2K warm-up iterations, and a minimum lr of 10−810^{-8}. Each sample contains 3–16 frames resized to 224 × 448. Training takes 40 hours on 2 RTX PRO 6000. Only the voxel-aligned feed-forward 3DGS module is optimized during training. The loss contains depth loss Ldepth\mathcal{L}_{\mathrm{depth}} and RGB loss Lrgb\mathcal{L}_{\mathrm{rgb}}, as shown in Equation (6).

L=Ldepth+Lrgb=λℓ1Ldepthℓ1+λalignLdepthalign+λℓ1Lrgbℓ1+λssimLrgbssim(6)\mathcal{L}=\mathcal{L}_{\mathrm{depth}}+\mathcal{L}_{\mathrm{rgb}}=\lambda_{\ell^1}\mathcal{L}_{\mathrm{depth}}^{\ell^1}+\lambda_{\mathrm{align}}\mathcal{L}_{\mathrm{depth}}^{\mathrm{align}}+\lambda_{\ell^1}\mathcal{L}_{\mathrm{rgb}}^{\ell^1}+\lambda_{\mathrm{ssim}}\mathcal{L}_{\mathrm{rgb}}^{\mathrm{ssim}} \tag*{(6)}

Here, Ldepthℓ1\mathcal{L}_{\mathrm{depth}}^{\ell^1} supervises rendered depth with ground-truth depth, whereas Ldepthalign\mathcal{L}_{\mathrm{depth}}^{\mathrm{align}} enforces alignment with VFM-predicted depth. Lrgbℓ1\mathcal{L}_{\mathrm{rgb}}^{\ell^1} and Lrgbssim\mathcal{L}_{\mathrm{rgb}}^{\mathrm{ssim}} are the ℓ1\ell_1 and SSIM [51] losses between rendered and ground-truth images. We set λℓ1\lambda_{\ell^1}, λalign\lambda_{\mathrm{align}}, λℓ1\lambda_{\ell^1}, and λssim\lambda_{\mathrm{ssim}} to 0.25, 0.25, 0.9, and 1.0.

Experiments

Baselines

We compare VoxelTTO with state-of-the-art (SOTA) baselines from two families: VGGT-style VFMs [44] and pixelSplat-style [5] feed-forward 3DGS. VFM baselines include AnySplat [19], YoNoSplat [56], DepthAnything3 (DA3) [27], and MapAnything (MA) [21]. AnySplat and DepthAnything3 support camera-parameter encoding. The initialization of vanilla 3DGS is used on MapAnything. [22] PixelSplat-style methods typically require camera parameters; we evaluate VolSplat [48], MVSplat [6] and LongLRM [62]. Experiments are conducted on the Replica [40], Tanks and Temples (TAT) [24], RE10K [61], and DTU [1] datasets, including indoor and outdoor, synthetic and real-world, scene-level and object-level scenarios. Per-scene results and additional visualizations are provided in the supplementary material.

RGB Novel-View Synthesis

Following prior works[22], we compare RGB NVS performance across VoxelTTO and the baselines. We use PSNR (P), SSIM (S) [51], and LPIPS (L) [59]. As shown in Table 1 and Figure 5, VoxelTTO performs strongly in both input and test views. TTO reduces pose-estimation error from the VFM backbone and substantially improves test-view NVS. Adding TTO also improves input-view NVS on Replica and RE10K, whereas leads to degradation in input-view on TAT despite improving test-view performance. This degradation may partly reflect inaccuracies in the ground-truth TAT camera poses.

Qualitative RGB novel-view synthesis results for test views

Figure 5. Qualitative RGB novel-view synthesis results for test views.

MethodReplica DatasetTAT DatasetRE10K Dataset
Input ViewsTest ViewsInput ViewsTest ViewsInput ViewsTest Views
P↑\uparrowS↑\uparrowL↓\downarrowP↑\uparrowS↑\uparrowL↓\downarrowP↑\uparrowS↑\uparrowL↓\downarrowP↑\uparrowS↑\uparrowL↓\downarrowP↑\uparrowS↑\uparrowL↓\downarrowP↑\uparrowS↑\uparrowL↓\downarrow
VolSplat16.150.6670.47115.440.6360.46715.450.5950.41916.330.7330.24120.550.7310.41514.130.5520.546
MVSplat19.970.7590.31819.410.7360.31916.480.6150.38015.050.5220.46014.720.6480.44114.640.6110.488
LongLRM27.940.8820.21223.390.8390.27622.010.7050.30212.180.4150.60026.120.8930.20617.440.8140.322
DA323.640.8370.20822.820.8170.24319.200.6210.34817.680.5560.41322.470.8180.23317.890.6940.355
VoxelTTO29.910.8960.17028.540.8760.17220.890.6560.34719.230.5650.39127.140.9010.20720.930.7640.279
AnySplat23.250.8250.20017.780.6810.35917.340.6450.31112.710.4400.47823.610.8420.21612.150.5330.534
YoNoSplat24.520.8090.26211.570.6350.43213.610.4770.51613.290.6600.30425.350.8370.1898.340.6010.663
MA14.350.5790.53317.950.7570.22712.660.4970.57014.660.5920.38113.760.5830.58910.010.4300.673
VoxelTTO (w/o TTO)29.280.8900.17425.830.8270.19623.740.7670.24714.710.4850.47526.970.8930.21617.460.6730.378

Table 1. Quantitative RGB novel-view synthesis results for input and test views.

Depth Novel-View Synthesis

To assess the geometric accuracy, we compare depth NVS performance with each baseline. We report the absolute error RMSE (cm), relative error Rel, and depth accuracy δτ\delta_{\tau}. V\mathcal{V} denote the set of pixels with valid ground-truth depth. For each pixel u∈Vu \in\mathcal{V}, d^u\hat{d}_{u} and dud_{u} denote the rendered and ground-truth depths, respectively. Rel(%) is used to evaluate relative depth error: Rel⁡=(1/∣V∣)∑u∈V(∣d^u−du∣/du)\operatorname{Rel} = (1/|\mathcal{V}|)\sum_{u \in\mathcal{V}}(|\hat{d}_{u} - d_{u}|/d_{u}). Depth accuracy δτ\delta_{\tau} (%) is the proportion of pixels in V\mathcal{V} satisfying max⁡(d^u/du,du/d^u)<τ\max(\hat{d}_{u}/d_{u},d_{u}/\hat{d}_{u}) < \tau. As shown in Figure 6 and Table 2, VoxelTTO’s geometric accuracy is higher than the baselines in both input and test views. TTO mitigates the misalignment caused by VFM and substantially improves test-view geometry.

Qualitative depth novel-view synthesis results for test views

Figure 6. Qualitative depth novel-view synthesis results for test views.

MethodInput ViewsTest Views
RMSE↓\downarrowRel↓\downarrowδ1.05↑\delta_{1.05}\uparrowδ1.10↑\delta_{1.10}\uparrowδ1.25↑\delta_{1.25}\uparrowRMSE↓\downarrowRel↓\downarrowδ1.05↑\delta_{1.05}\uparrowδ1.10↑\delta_{1.10}\uparrowδ1.25↑\delta_{1.25}\uparrow
Replica Dataset
VolSplat102.830.4211.8822.5546.78202.263.890.8122.4436.859
MVSplat216.069.261.9843.8949.2611649446.111.8425.7742.37
LongLRM18.066.01860.7383.0996.8319.005.72465.8785.5297.14
DA3125.543.2728.3046.4057.85172.038.6964.1871.1976.29
VoxelTTO6.1670.68898.6699.2999.726.9450.86698.8699.2499.58
AnySplat33.729.58259.0576.8886.1435.9210.7845.4367.0989.86
YoNoSplat23.477.28151.0775.5894.78242.178.730.3040.6922.643
MA13.303.55982.6993.7498.1237.9610.1544.8068.3786.92
VoxelTTO (w/o TTO)7.3940.77798.5799.2299.669.0321.11598.1498.8399.38
TAT Dataset
VolSplat490.733.739.98419.1341.2569.1531.059.05818.0142.70
MVSplat894.371.980.9111.8094.739143.776.570.3430.6992.203
LongLRM181.624.674.7579.94054.43462.116.3442.0859.3780.53
DA3456.941.4418.7636.7160.34299.430.493.9348.18123.29
VoxelTTO90.632.87590.1295.8398.1546.223.23586.7595.1698.33
AnySplat117.44.79370.0190.9896.95499.1360.35.40515.3730.40
YoNoSplat511.537.068.01815.6933.5483.5747.597.48015.3838.70
MA139.96.56067.0181.1194.77512.4399.115.2828.0631.31
VoxelTTO (w/o TTO)82.072.07996.3597.4598.4963.937.03552.3978.8495.44

Table 2. Quantitative depth novel-view synthesis results for input and test views.

Pose Alignment Estimation

We use absolute translation error and rotation error to evaluate the pose alignment. Translation error is the Euclidean distance between predicted and ground truth. Rotation error is the geodesic distance on the rotation manifold between prediction and ground truth: eR=arccos⁡((tr⁡((Rip)TRig)−1)/2)e_{R} = \arccos((\operatorname{tr}((R_{i}^{p})^{\mathsf{T}}R_{i}^{g}) - 1)/2). As shown in Table 3, VoxelTTO freezes both the VFM backbone and pose decoder. Therefore, removing TTO yields pose alignment accuracy comparable to DA3. The DA3 result in Table 3 uses encoded camera parameters and is consequently slightly improved. Even without TTO, VoxelTTO achieves stronger appearance and geometric performance, as discussed above. Adding TTO substantially improves alignment between predicted and ground-truth poses.

MethodReplicaTAT
Rot(°)↓Trans(m)↓Rot(°)↓Trans(m)↓
DA30.3240.0040.6820.030
VoxelTTO0.2200.0020.2620.020
AnySplat7.3200.1063.5160.135
YoNoSplat4.7690.0635.3180.115
MA4.4380.0563.2320.065
VoxelTTO (w/o TTO)0.3010.0061.6360.047

Table 3. Mean input-camera alignment errors.

Surface Alignment

As coupled errors in VFM poses and depth maps can displace the predicted point cloud from the ground-truth surface, further validation is carried out on DTU. We compare VoxelTTO without TTO, VoxelTTO with DA3 camera-parameter encoding [27] instead of TTO, and the full VoxelTTO model. Following DTU[1], we report accuracy (Acc.), completeness (Com.), and their mean (Over.), which are all computed with Chamfer distance. As shown in Table 4, camera-parameter encoding suppresses misalignment to some extent. VoxelTTO further reduces accuracy, completeness, and overall distance.

MethodAcc.↓Com.↓Over.↓
VoxelTTO (w/o TTO)5.5569.1597.357
VoxelTTO (Camera Enc.)3.7194.8534.286
VoxelTTO2.9523.8803.416

Table 4. Chamfer distances for surface alignment on DTU.

Ablation Studies

All the ablation studies are shown in Table 5. For the self-splitting Gaussian decoder, we evaluate the number of Gaussian splits per voxel. Increasing the number of splits improves geometry and appearance, but the improvement from 2 to 4 splits is no longer substantial. To evaluate the effect of stochastic solid volume rendering on geometry, we compare it with 3DGS rasterization during both training and inference. Using 3DGS rasterization substantially reduces appearance and geometric accuracy under both stages. Here, V-R denotes stochastic solid volume rendering supervision (V) during training and 3DGS rasterization (R) for NVS at inference. We compare different voxel sizes in voxel-aligned 3DGS. Both appearance and geometric performance decline as voxel size increases. Balancing performance and computation, and following prior work [19], we set the size to 0.002m. An RGB loss LrgbL_{\mathrm{rgb}} is required to regress color. We therefore ablate the depth loss separately. Removing the depth loss slightly reduces geometric accuracy. We further evaluate the effect of TTO iterations. As shown in Figure 7, test-view appearance performance (LPIPS) and geometric performance (Rel) improve substantially as the number of TTO iterations increases. We select 20 iterations by balancing accuracy and efficiency.

Test-view RGB-D synthesis performance at different TTO iteration counts

Figure 7. Test-view RGB-D synthesis performance at different TTO iteration counts.

ExperimentSettingPSNR↑\uparrowSSIM↑\uparrowLPIPS↓\downarrowRMSE (cm)↓\downarrowRel↓\downarrowδ1.05↑\delta_{1.05}\uparrowδ1.10↑\delta_{1.10}\uparrowδ1.25↑\delta_{1.25}\uparrow
Number of Gaussian Splits127.820.8660.1728.090.9498.7199.1099.47
Number of Gaussian Splits2 (Ours)28.54<u>0.876</u><u>0.172</u>6.950.87<u>98.86</u><u>99.24</u>99.58
Number of Gaussian Splits4<u>28.35</u>0.8760.179<u>7.21</u>0.8698.9399.2499.54
Stochastic Solid Volume RenderingV-V (Ours)28.540.8760.1726.950.8798.8699.2499.58
Stochastic Solid Volume RenderingV-R<u>27.11</u><u>0.865</u><u>0.188</u><u>12.01</u><u>1.24</u><u>97.96</u><u>98.62</u>99.15
Stochastic Solid Volume RenderingR-R21.520.7740.26927.584.0587.7695.1197.10
Stochastic Solid Volume RenderingR-V22.370.7980.26015.823.0489.3297.0398.66
Voxel Size0.001529.340.8910.1426.620.8399.0499.3699.64
Voxel Size0.002 (Ours)28.54<u>0.876</u><u>0.172</u><u>6.95</u><u>0.87</u><u>98.86</u><u>99.24</u><u>99.58</u>
Voxel Size0.00326.780.8560.2119.451.1898.3298.8899.32
Voxel Size0.00523.440.7980.29015.222.4593.1797.1598.47
Depth LossOurs28.540.876<u>0.172</u>6.950.8798.8699.2499.58
Depth LossNo LdepthL_{\mathrm{depth}}28.25<u>0.873</u>0.171<u>7.49</u>0.94<u>98.67</u><u>99.10</u><u>99.53</u>

Table 5. Ablation studies on test-view RGB-D NVS on the Replica dataset.

Runtime

For ten 518×518518 \times518 images, VoxelTTO without TTO takes 2.304 s, compared with 2.293 s for DA3, indicating negligible overhead from voxel decoding, as shown in Figure 8. We retain frozen attention computations in BF16 and LoRA computations in Float32, while baselines use their official precision. BF16 reduces backbone runtime from 1.959 s

Runtime of individual modules in VoxelTTO without TTO and in the baselines

Figure 8. Runtime of individual modules in VoxelTTO without TTO and in the baselines.

to 0.1956 s. Each TTO step requires 0.266 s forward and 0.410 s backward, so 20 steps complete in under 14 s and total runtime is under 14.4 s.

Conclusion

We presented VoxelTTO, a voxel-aligned feed-forward 3DGS framework with test-time optimization. First, it reduces the excessive overlap and artifacts of pixel-aligned feed-forward 3DGS by aggregating dense pixel features into global voxel features. A sparse 3D voxel U-Net and self-splitting decoder regress Gaussian attributes. Stochastic solid volume rendering replaces 3DGS rasterization during training and inference, improving the geometric accuracy of the Gaussian scene. Second, pose-supervised TTO adapts lightweight LoRA modules in the VFM backbone to align predicted camera parameters with the provided camera priors. Extensive experiments on DTU, Replica, RE10K, and TAT show that VoxelTTO reconstructs geometrically accurate 3DGS scenes and achieves SOTA test-view RGB and depth rendering accuracy.

Reproducibility statement

The train dataset: ScanNet++[57]https://scannetpp.mlsg.cit.tum.de/scannetpp; Infinigen[34]https://github.com/princeton-vl/infinigen; ARKitScenes[4]https://github.com/apple/aiml-research/ARKitScenes; DL3DV[28]https://dl3dv-10k.github.io/DL3DV-10K/

The eval dataset: Replica[40]https://github.com/facebookresearch/replica-dataset; Tanks and Temples[24]https://www.tanksandtemples.org/; DTU[1]http://roboimagedata.compute.dtu.dk; RE10K[61]https://huggingface.co/datasets/HyelinKim/RE10K-test

The code: https://github.com/Zhaoyibinn/OctreeSplat

LLM Usage Statement

The authors used ChatGPT to assist with code implementation and language polishing to improve the readability of the manuscript. The authors verified the resulting content and take full responsibility for the paper.

Appendix

Model Structure

VoxelTTO uses the DA3-GIANT backbone initialized with pretrained weights [27]. The backbone contains 40 Transformer blocks with a token dimension of 1536 and a patch size of 14×1414 \times14. The pretrained depth-confidence head and camera decoder are retained. The pretrained backbone and prediction heads remain frozen during VoxelTTO’s training.

A separate trainable feature DPT uses the same backbone outputs (fused by layers 20, 28, 34, 40) and feature channels (256, 512, 1024, 1024) as the depth head. The final prediction layer is removed, yielding a dense 128-channel feature map with the same resolution as the input image. Predicted depth and camera parameters are used to unproject into a global coordinate system. At each pixel, the resulting 3D point is associated with its dense feature, RGB color, and depth confidence. We aggregate points into sparse voxels using confidence-weighted averages for their coordinates, features, and colors. Negative confidence weights are clamped to zero and deleted. The minimum point confidence in each voxel is retained for Gaussian decoding.

The voxel encoder is a sparse 3D U-Net implemented with spconv. A 5×5×55 \times5 \times5 submanifold convolution maps the 128-channel input to 32 channels. Four encoder stages use [32, 64, 128, 256] channels and [2, 3, 4, 6] residual blocks. Each stage starts with a 2 stride sparse convolution with a 2×2×22 \times2 \times2 kernel. Four decoder stages ordered from coarse to fine, using [256, 128, 96, 96] channels and 2 residual blocks per stage. Sparse inverse convolutions upsample the features, which are concatenated with the corresponding encoder. Each residual block contains two 3×3×33 \times3 \times3 submanifold convolutions, combined with BatchNorm and ReLU. The 1×1×11 \times1 \times1 submanifold convolution produces a 128-dimensional feature for each input voxel.

Each voxel feature is processed by separate geometry and appearance MLPs. The structure is linear layers with [128, 256, 128]-dimensional features and GELUs. The geometry branch predicts primitive center-offset parameters zv,jz_{v,j}, scales sv,js_{v,j}, rotation quaternions rv,jr_{v,j}, and opacity logits ov,jo_{v,j}; the appearance branch predicts 2-degree spherical harmonics (SH) coefficients. As each voxel splits into nn (default as 2) Gaussians, the output dimensions are 3n3n, 3n3n, 4n4n, nn, and 27n27n. For voxel anchor xvx_v, the primitive’s center is calculated

is calculated as shown in Equation (7).

μv,j=xv+Δmax⁡[sigmoid⁡(gzv,j)−12],(7)\mu_{v,j} = x_v + \Delta_{\max}\left[\operatorname{sigmoid}(g_{z_{v,j}}) - \frac{1}{2}\right], \tag*{(7)}

where zv,jz_{v,j} is the center-offset head output, qq is a learned scalar initialized to 10−410^{-4}, and Δmax⁡=0.004\Delta_{\max} = 0.004. The center head is zero-initialized, so all Gaussians initially lie at the voxel anchor. Scales sv,js_{v,j} are mapped by a sigmoid to [10−5,0.002][10^{-5}, 0.002] in the voxel coordinate system. Quaternions rv,jr_{v,j} are normalized to unit length. Voxel confidence κv\kappa_v adjusts each opacity logit ov,jo_{v,j} as shown in Equation (8).

σv,j=sigmoid⁡(ov,j+clip⁡(κv−5,−5,5)).(8)\sigma_{v,j} = \operatorname{sigmoid}(o_{v,j} + \operatorname{clip}(\kappa_v - 5, -5, 5)). \tag*{(8)}

The constant SH coefficients are initialized with the voxel RGB color by (av−0.5)/0.28209479(a_v - 0.5)/0.28209479, where the predicted SH residual added. The SH head is also zero-initialized. Following DA3, coefficients at degrees 1 and 2 are multiplied by 0.025 and 0.00625 to attenuate higher-order appearance residuals.

Enabling BF16 converts the frozen Transformer, the feature projection, and fusion layers of DPTs to BF16. The final depth and confidence prediction layers, camera encoder and decoder, sparse voxel U-Net, and Gaussian primitive decoder retain FP32 parameters. LoRA matrices are also stored in FP32. Gaussian scene parameters are represented in FP32 for rendering.

During computation, Transformer linear operations and DPT feature projection and fusion use BF16. Dense feature maps, cached Transformer tokens, and aggregated voxel features also use BF16. Features are explicitly converted to FP32 before the final depth and confidence prediction layers and the sparse U-Net. Camera decoding, depth un-projection, sparse voxel convolutions, Gaussian attribute decoding, and rendering operate on FP32. Voxelization uses FP32 point coordinates, while feature aggregation uses BF16. During TTO, the frozen Transformer retains the same BF16 parameters and activations. Each LoRA branch casts inputs to FP32 and performs calculations in FP32. The residual is then cast to BF16 before adding to the linear layer. Backpropagation traverses the mixed-precision computation graph, while LoRA parameter gradients and Adam optimizer states remain in FP32.

Scale ambiguity

Global Sim(3) alignment is performed after scene prediction, including TTO when enabled, and before rendering and evaluation in the ground-truth coordinate frame. We estimate this transform from the input cameras and apply it to the predicted cameras and reconstructed scene. Image-based reconstruction inevitably suffers from scale ambiguity. In VoxelTTO, we explicitly construct voxels. The default voxel size (0.002 m) is specified in DA3’s[27] estimated metric scale. DA3 converts its raw depth prediction to metric depth as dmetric=f⋅dnet/300d_{\mathrm{metric}} = f \cdot d_{\mathrm{net}}/300, where dnetd_{\mathrm{net}} is the network output and ff is the focal length predicted in pixels. Because this metric scale is estimated, subsequent Sim(3) alignment with the ground-truth cameras may introduce a scale correction. If the alignment scale is saligns_{\mathrm{align}}, the voxel size expressed in the ground-truth coordinate system becomes ∣salign∣×0.002\lvert s_{\mathrm{align}}\rvert\times0.002 m. Thus, voxels’ final physical size may differ slightly from 0.002 m, depending on the accuracy of DA3’s metric-scale estimate.

Sim(3) alignment. We estimate a similarity transform T\mathcal{T} between predicted and ground-truth input-camera centers using Umeyama alignment [41]. For world-to-camera extrinsics Ei=[Ri∣ti]\mathbf{E}_i = [\mathbf{R}_i \mid\mathbf{t}_i], the camera center is ci=−RiTti\mathbf{c}_i = -\mathbf{R}_i^{\mathsf{T}}\mathbf{t}_i. We parameterize the transform as T(x)=salignQx+b\mathcal{T}(\mathbf{x}) = s_{\mathrm{align}}\mathbf{Q}\mathbf{x} + \mathbf{b}, where salign>0s_{\mathrm{align}} > 0 and Q∈SO(3)\mathbf{Q} \in\mathrm{SO}(3). The optimal transform T∗\mathcal{T}^{*} maps the prediction frame to the ground-truth frame, as shown in Equation (9).

T∗=arg⁡min⁡T∈Sim(3) 1M∑i=1M∥cig−T(cip)∥22.(9)\mathcal{T}^{*} = \underset{\mathcal{T}\in\mathrm{Sim}(3)}{\arg\min}\ \frac{1}{M}\sum_{i=1}^{M}\left\lVert\mathbf{c}^{g}_{i} - \mathcal{T}\left(\mathbf{c}^{p}_{i}\right)\right\rVert_{2}^{2}. \tag*{(9)}

Here, pp and gg denote predicted and ground-truth quantities. We define their cross-covariance matrix H\mathbf{H} and the reflection-correction matrix J\mathbf{J} as shown in Equation (10).

H=1M∑i(cig−cˉg)(cip−cˉp)T=UΛVT,\mathbf{H} = \frac{1}{M}\sum_i\left(\mathbf{c}^{g}_{i}-\bar{\mathbf{c}}^{g}\right)\left(\mathbf{c}^{p}_{i}-\bar{\mathbf{c}}^{p}\right)^{\mathsf{T}} = \mathbf{U}\boldsymbol{\Lambda}\mathbf{V}^{\mathsf{T}},
J=diag⁡(1,1,det⁡(UVT)).(10)\mathbf{J} = \operatorname{diag}\left(1,1,\det\left(\mathbf{U}\mathbf{V}^{\mathsf{T}}\right)\right). \tag*{(10)}

Here, UΛVT\mathbf{U}\boldsymbol{\Lambda}\mathbf{V}^{\mathsf{T}} is the singular value decomposition of H\mathbf{H}. The matrix J\mathbf{J} ensures that the recovered rotation has determinant +1+1. The parameters (salign,Q,b)(s_{\mathrm{align}},\mathbf{Q},\mathbf{b}) of T∗\mathcal{T}^{*} have the closed-form solution, as shown in Equation (11).

Q=UJVT,\mathbf{Q} = \mathbf{U}\mathbf{J}\mathbf{V}^{\mathsf{T}},
salign=tr⁡(ΛJ)M−1∑i∥cip−cˉp∥22,s_{\mathrm{align}} = \frac{\operatorname{tr}(\boldsymbol{\Lambda}\mathbf{J})}{M^{-1}\sum_i\left\lVert\mathbf{c}^{p}_{i}-\bar{\mathbf{c}}^{p}\right\rVert_{2}^{2}},
b=cˉg−salignQcˉp.(11)\mathbf{b} = \bar{\mathbf{c}}^{g} - s_{\mathrm{align}}\mathbf{Q}\bar{\mathbf{c}}^{p}. \tag*{(11)}

The same transform T∗\mathcal{T}^{*} is applied to the reconstructed scene and predicted cameras. For a Gaussian center μq\boldsymbol{\mu}_q, scale vector sq\mathbf{s}_q, and rotation matrix Rq\mathbf{R}_q, the transform is shown in in Equation (12).

μq′=T∗(μq)=salignQμq+b,\boldsymbol{\mu}'_q = \mathcal{T}^{*}(\boldsymbol{\mu}_q) = s_{\mathrm{align}}\mathbf{Q}\boldsymbol{\mu}_q + \mathbf{b},
sq′=salignsq,Rq′=QRq.(12)\mathbf{s}'_q = s_{\mathrm{align}}\mathbf{s}_q,\qquad\mathbf{R}'_q = \mathbf{Q}\mathbf{R}_q. \tag*{(12)}

We also rotate the SH coefficients consistently with Q\mathbf{Q}, as shown in Equation (13).

hq,ℓ′(k)=D(ℓ)(Q)hq,ℓ(k),ℓ∈{0,1,2},h_{q,\ell}^{\prime(k)} = \mathbf{D}^{(\ell)}(\mathbf{Q})h_{q,\ell}^{(k)},\qquad\ell\in\{0,1,2\},
hq,ℓm′(k)=∑m′=−ℓℓDmm′(ℓ)(Q)hq,ℓm′(k).(13)h_{q,\ell m}^{\prime(k)} = \sum_{m'=-\ell}^{\ell}D_{mm'}^{(\ell)}(\mathbf{Q})h_{q,\ell m'}^{(k)}. \tag*{(13)}

Here, D(ℓ)(Q)\mathbf{D}^{(\ell)}(\mathbf{Q}) is the real Wigner rotation matrix expressed in the renderer’s SH basis convention. The degree-zero coefficient remains unchanged.

TTO loss uses a separate normalization to remove global frame and scale ambiguity. For each pose set, we invert the extrinsics to obtain camera-to-world poses Ti=Ei−1\mathbf{T}_i = \mathbf{E}_i^{-1} and compute T~i=T1−1Ti\widetilde{\mathbf{T}}_i = \mathbf{T}_1^{-1}\mathbf{T}_i, to be expressed in respective first-camera coordinate. t~i\widetilde{\mathbf{t}}_i is the translation of T~i\widetilde{\mathbf{T}}_i, the calculation of normalized translations is shown in Equation (14).

t^i=t~imax⁡(M−1∑j∥t~j∥2,10−6).(14)\widehat{\mathbf{t}}_i = \frac{\widetilde{\mathbf{t}}_i}{\max\left(M^{-1}\sum_j\left\lVert\widetilde{\mathbf{t}}_j\right\rVert_2,10^{-6}\right)}. \tag*{(14)}

This normalization is computed before applying the rotation and translation TTO losses. The translation loss is invariant to global scale after normalization. Intrinsics are also normalized as (fx/W,fy/H,cx/W,cy/H)(f_x/W,f_y/H,c_x/W,c_y/H), independently of camera resolution and scale.

Detailed experimental results

For evaluation on Replica [40], we follow the experimental protocol of GS-ICP-SLAM [14] and evaluate 8 scenes. We uniformly sample 10 input images per scene at intervals of 50 frames. Test images are sampled every 5 frames within the same frame range, excluding all input images. All images are center-cropped and resized to 518×518518 \times518 pixels. The same image preprocessing is applied to all other datasets. Detailed RGB view synthesis results on Replica dataset are shown in Table A1 and Table A2. Detailed depth view synthesis results on Replica dataset are shown in Table A3 and Table A4. For all baselines, both depth and RGB images are rendered using their respective default renderer.

MethodMetricoffice0office1office2office3office4room0room1room2Average
VolSplatPSNR↑17.6026.4612.3012.9813.9615.5315.2215.1416.15
VolSplatSSIM↑0.6940.8310.5280.6110.6790.6540.6610.6800.667
VolSplatLPIPS↓0.4360.2850.5780.4910.4830.5160.4790.4990.471
MVSplatPSNR↑21.4624.8317.6617.8517.6720.5519.3220.4019.97
MVSplatSSIM↑0.7920.7820.6670.7300.7740.8120.7470.7680.759
MVSplatLPIPS↓0.3290.2240.3620.3150.3540.2800.3250.3560.318
AnySplatPSNR↑26.4027.5719.4519.8724.4022.8921.7923.5923.25
AnySplatSSIM↑0.8740.8380.7570.7770.8660.8360.8310.8240.825
AnySplatLPIPS↓0.1650.1940.2270.2430.1920.1930.1740.2150.200
YoNoSplatPSNR↑25.1229.1222.0924.3625.9122.8322.2424.4924.52
YoNoSplatSSIM↑0.8070.8820.7530.8230.8520.8050.7500.8030.809
YoNoSplatLPIPS↓0.2850.2270.2760.2150.2300.2600.3140.2850.262
DepthAnything3PSNR↑24.5128.6621.8022.4024.2723.7320.7423.0123.64
DepthAnything3SSIM↑0.8310.8600.8120.8360.8780.8620.7610.8520.837
DepthAnything3LPIPS↓0.2360.1490.2240.1830.1730.2030.2820.2170.208
MapAnythingPSNR↑16.8824.3712.0012.8713.0911.4711.0013.0714.35
MapAnythingSSIM↑0.6410.7130.5010.5680.5980.5420.4950.5710.579
MapAnythingLPIPS↓0.4940.3490.5720.5520.5490.5660.6260.5570.533
LongLRMPSNR↑26.2029.3927.3227.6329.3528.4526.0229.1527.94
LongLRMSSIM↑0.8520.8450.8630.8910.9230.9140.8640.9040.882
LongLRMLPIPS↓0.2540.2680.1910.1790.1700.1930.2360.2070.212
VoxelTTO (w/o TTO)PSNR↑33.4533.1025.8427.4229.8327.9827.8128.8229.28
VoxelTTO (w/o TTO)SSIM↑0.9210.9380.8490.8870.9030.8960.8420.8820.890
VoxelTTO (w/o TTO)LPIPS↓0.1380.1550.1880.1430.1800.1820.2040.1990.174
OursPSNR↑33.4432.7827.2026.9730.6728.8828.9030.4729.91
OursSSIM↑0.9200.9300.8660.8790.9130.9070.8500.9010.896
OursLPIPS↓0.1420.1620.1830.1580.1700.1710.1920.1860.170

Table A1. Per-scene RGB view synthesis results on Replica input views.

MethodMetricoffice0office1office2office3office4room0room1room2Average
VolSplatPSNR↑16.7026.1311.6711.5013.0614.8814.6914.8715.44
VolSplatSSIM↑0.6400.8570.4880.5380.6320.6240.6570.6530.636
VolSplatLPIPS↓0.4500.2490.5790.5010.5130.4930.4530.4960.467
MVSplatPSNR↑20.5123.6116.7417.6116.4620.9919.8819.5019.41
MVSplatSSIM↑0.7660.7400.6400.6950.7470.8020.7670.7310.736
MVSplatLPIPS↓0.2720.2540.3700.3360.3870.2930.2490.3920.319
AnySplatPSNR↑15.9922.7415.3115.3818.4619.7516.4618.1917.78
AnySplatSSIM↑0.6660.7320.5950.6030.7500.7310.6860.6860.681
AnySplatLPIPS↓0.3870.3250.4170.3830.3260.3100.3550.3690.359
YoNoSplatPSNR↑12.1117.438.719.3310.6411.1710.9412.2411.57
YoNoSplatSSIM↑0.5970.5300.5940.6180.6910.7000.6610.6890.635
YoNoSplatLPIPS↓0.4070.3860.4860.4550.3750.4170.4740.4530.432
DepthAnything3PSNR↑24.4127.6720.2921.7222.9223.7619.2522.5722.82
DepthAnything3SSIM↑0.8000.8370.7880.8240.8560.8630.7140.8490.817
DepthAnything3LPIPS↓0.2500.2050.2560.2240.2200.2380.3120.2420.243
MapAnythingPSNR↑20.6924.3715.3714.8115.5317.2917.0918.4917.95
MapAnythingSSIM↑0.8100.7720.6880.6820.7530.7780.7990.7760.757
MapAnythingLPIPS↓0.1950.2670.2540.2540.2620.2080.1670.2120.227
LongLRMPSNR↑27.0329.0321.9621.4622.7822.4817.8324.5523.39
LongLRMSSIM↑0.8440.8920.8160.8260.8750.8640.7410.8570.839
LongLRMLPIPS↓0.2890.2680.2810.2490.2480.2570.3480.2650.276
VoxelTTO (w/o TTO)PSNR↑26.3029.1020.8925.4129.0925.8423.5926.4625.83
VoxelTTO (w/o TTO)SSIM↑0.7990.8830.7370.8510.8990.8630.7420.8380.827
VoxelTTO (w/o TTO)LPIPS↓0.2090.1790.2320.1480.1810.1910.2230.2040.196
OursPSNR↑29.5627.9924.8727.9030.7528.8828.2430.1428.54
OursSSIM↑0.8570.8760.8180.8920.9200.9050.8330.9100.876
OursLPIPS↓0.1680.1880.1980.1480.1660.1670.1710.1690.172

Table A2. Per-scene RGB view synthesis results on Replica test views.

MethodMetricoffice0office1office2office3office4room0room1room2Average
VolSplatMAE (cm)↓\downarrow99.5222.5691.10104.198.4284.8696.1771.3883.52
VolSplatRMSE (cm)↓\downarrow122.227.92115.9128.9117.5109.2115.885.66102.8
VolSplatRel (%)↓\downarrow41.2312.3828.1430.7034.1824.2541.0031.5330.42
VolSplatδ1.05\delta_{1.05} (%)↑\uparrow8.30125.9311.249.2678.67414.178.7888.68711.88
VolSplatδ1.10\delta_{1.10} (%)↑\uparrow15.5848.3621.8918.0616.5725.8416.7317.4422.55
VolSplatδ1.25\delta_{1.25} (%)↑\uparrow32.3985.1046.7041.7836.5055.3935.1041.3646.78
MVSplatMAE (cm)↓\downarrow133.1146.1281.7223.5246.9138.8155.6195.8190.2
MVSplatRMSE (cm)↓\downarrow151.3161.3303.8263.4275.6184.6172.9215.7216.0
MVSplatRel (%)↓\downarrow56.7572.5686.4763.6981.5842.9067.6182.5969.26
MVSplatδ1.05\delta_{1.05} (%)↑\uparrow4.3692.7380.62781.9620.74222.5272.0380.87121.984
MVSplatδ1.10\delta_{1.10} (%)↑\uparrow8.5975.2291.2253.8471.5065.0284.0311.6893.894
MVSplatδ1.25\delta_{1.25} (%)↑\uparrow18.9711.462.8869.2343.85213.4710.214.0189.261
AnySplatMAE (cm)↓\downarrow15.3312.4916.7210.6613.087.317102.710.3523.58
AnySplatRMSE (cm)↓\downarrow21.2415.1424.4718.0023.9012.73133.520.9233.72
AnySplatRel (%)↓\downarrow6.7116.8575.3493.1694.6972.29343.264.3199.582
AnySplatδ1.05\delta_{1.05} (%)↑\uparrow48.6647.2753.3684.0367.7592.500.0553978.8159.05
AnySplatδ1.10\delta_{1.10} (%)↑\uparrow73.8174.4791.5895.3488.4797.230.120794.0276.88
AnySplatδ1.25\delta_{1.25} (%)↑\uparrow98.6299.1098.0098.7298.3798.700.514697.1186.14
YoNoSplatMAE (cm)↓\downarrow31.2916.7413.2714.7818.9612.2819.5713.6317.56
YoNoSplatRMSE (cm)↓\downarrow37.7320.7721.0319.4226.0717.6025.0420.1323.47
YoNoSplatRel (%)↓\downarrow14.278.7954.3014.6796.8144.0198.8816.4947.281
YoNoSplatδ1.05\delta_{1.05} (%)↑\uparrow20.8835.4275.5163.5750.4473.4635.0854.2451.07
YoNoSplatδ1.10\delta_{1.10} (%)↑\uparrow40.7661.9090.2191.9683.1291.3563.7881.5975.58
YoNoSplatδ1.25\delta_{1.25} (%)↑\uparrow79.7394.1798.0899.3196.5999.1194.7496.5394.78
DepthAnything3MAE (cm)↓\downarrow232.926.7823.22346.419.0415.76232.120.04114.5
DepthAnything3RMSE (cm)↓\downarrow236.934.9842.54361.329.9026.60240.131.98125.5
DepthAnything3Rel (%)↓\downarrow100.0014.228.333100.007.4365.337100.010.9143.27
DepthAnything3δ1.05\delta_{1.05} (%)↑\uparrow021.5244.520.00417453.2572.70034.4728.30
DepthAnything3δ1.10\delta_{1.10} (%)↑\uparrow035.0480.940.00417484.8190.68079.7546.40
DepthAnything3δ1.25\delta_{1.25} (%)↑\uparrow085.1693.900.00417494.8596.98091.9657.85
MapAnythingMAE (cm)↓\downarrow8.2064.7668.6969.2168.3356.6388.6478.7727.909
MapAnythingRMSE (cm)↓\downarrow11.086.84915.5914.9617.4812.2512.4515.7613.30
MapAnythingRel (%)↓\downarrow3.6042.6123.0863.3733.7852.6574.1065.2543.559
MapAnythingδ1.05\delta_{1.05} (%)↑\uparrow75.8984.2486.1384.8687.6891.8668.6482.2582.69
MapAnythingδ1.10\delta_{1.10} (%)↑\uparrow95.3197.3693.0894.9392.4396.0492.9887.8693.74
MapAnythingδ1.25\delta_{1.25} (%)↑\uparrow99.7299.7498.6398.3696.3598.7899.1194.2998.12
LongLRMMAE (cm)↓\downarrow14.7311.7012.9918.6711.5810.6317.6211.2713.65
LongLRMRMSE (cm)↓\downarrow18.3014.6417.9923.0317.2413.9423.7515.6118.06
LongLRMRel (%)↓\downarrow6.5686.1904.4856.3524.6183.6839.0017.2536.018
LongLRMδ1.05\delta_{1.05} (%)↑\uparrow43.6850.1471.0052.5376.5977.1845.8268.9560.73
LongLRMδ1.10\delta_{1.10} (%)↑\uparrow76.2879.6289.9882.6089.4093.6771.0182.2483.09
LongLRMδ1.25\delta_{1.25} (%)↑\uparrow99.1898.5398.8198.1297.5899.3191.3691.7796.83
VoxelTTO (w/o TTO)MAE (cm)↓\downarrow1.3531.5302.7282.7282.5512.0751.7791.9692.089
VoxelTTO (w/o TTO)RMSE (cm)↓\downarrow4.0668.7879.5996.77310.645.2224.7829.2927.394
VoxelTTO (w/o TTO)Rel (%)↓\downarrow0.58370.83160.85060.84950.86320.62810.79870.81690.777
VoxelTTO (w/o TTO)δ1.05\delta_{1.05} (%)↑\uparrow98.7798.7997.9298.3498.7199.0399.1897.8798.57
VoxelTTO (w/o TTO)δ1.10\delta_{1.10} (%)↑\uparrow99.4499.4298.7999.1799.0899.5099.6198.7899.22
VoxelTTO (w/o TTO)δ1.25\delta_{1.25} (%)↑\uparrow99.8599.9299.5299.6999.4199.8199.8399.2999.66
OursMAE (cm)↓\downarrow1.4201.0402.5642.5382.0931.8731.7061.4931.841
OursRMSE (cm)↓\downarrow4.3923.0149.2786.4688.6445.4194.6407.4846.167
OursRel (%)↓\downarrow0.61030.54250.81110.80960.71500.58020.80500.63740.688
Oursδ1.05\delta_{1.05} (%)↑\uparrow98.7698.7897.9998.1899.0699.0599.2298.3198.66
Oursδ1.10\delta_{1.10} (%)↑\uparrow99.4199.3798.8699.2099.3599.4799.6299.0799.29
Oursδ1.25\delta_{1.25} (%)↑\uparrow99.8899.9299.5899.7099.6099.7999.8399.4899.72

Table A3. Per-scene depth view synthesis results on Replica input views.

MethodMetricoffice0office1office2office3office4room0room1room2Average
VolSplatMAE (cm)↓\downarrow168.245.77225.5259.5209.1247.9187.0175.8189.9
VolSplatRMSE (cm)↓\downarrow175.552.01242.9275.1226.3260.0196.3190.3202.2
VolSplatRel (%)↓\downarrow70.9622.5269.4269.9865.0369.3875.7168.1563.89
VolSplatδ1.05\delta_{1.05} (%)↑\uparrow0.24595.6060.0017000.012070.13460.0015240.046090.44910.812
VolSplatδ1.10\delta_{1.10} (%)↑\uparrow0.667817.010.019270.051750.68360.015540.20280.89182.443
VolSplatδ1.25\delta_{1.25} (%)↑\uparrow2.15544.840.20420.17294.0290.21571.1012.1536.859
MVSplatMAE (cm)↓\downarrow29821531278.3166.1288.671.781951249.9939.7
MVSplatRMSE (cm)↓\downarrow42843178787.3370.5681.0175.93139584.81649
MVSplatRel (%)↓\downarrow1372847.4103.275.61135.824.46872.5138.3446.1
MVSplatδ1.05\delta_{1.05} (%)↑\uparrow0.406416.6514.2114.505.53927.824.69510.9011.84
MVSplatδ1.10\delta_{1.10} (%)↑\uparrow0.916035.4132.4035.4316.3851.728.94225.0325.77
MVSplatδ1.25\delta_{1.25} (%)↑\uparrow2.85148.7259.1160.9930.4575.0919.1342.6742.37
AnySplatMAE (cm)↓\downarrow42.2317.0219.6222.9417.5714.2045.1121.9125.08
AnySplatRMSE (cm)↓\downarrow51.2722.4333.7134.6230.7020.5255.3338.8035.92
AnySplatRel (%)↓\downarrow18.558.2427.1378.3786.8894.80320.9611.3210.78
AnySplatδ1.05\delta_{1.05} (%)↑\uparrow15.9032.4654.7062.9260.7879.7110.9346.1045.43
AnySplatδ1.10\delta_{1.10} (%)↑\uparrow31.2164.1879.3076.7690.3994.8325.3774.7467.09
AnySplatδ1.25\delta_{1.25} (%)↑\uparrow71.4694.9693.3994.0297.6597.4579.1490.8589.86
YoNoSplatMAE (cm)↓\downarrow173.0153.7262.2266.0248.3293.8212.4205.8226.9
YoNoSplatRMSE (cm)↓\downarrow180.5160.1282.2292.6267.6309.2222.7222.5242.1
YoNoSplatRel (%)↓\downarrow72.7376.5178.3074.4176.0983.2587.9480.6278.73
YoNoSplatδ1.05\delta_{1.05} (%)↑\uparrow1.8070.023970.059950.26660.045170.0174800.21140.304
YoNoSplatδ1.10\delta_{1.10} (%)↑\uparrow3.4860.056900.26731.0190.20360.060270.0014220.44770.692
YoNoSplatδ1.25\delta_{1.25} (%)↑\uparrow6.8910.39923.3606.0291.5670.81870.26421.8162.643
DepthAnything3MAE (cm)↓\downarrow291.639.5210.1916.8712.666.408375.910.9595.51
DepthAnything3RMSE (cm)↓\downarrow539.370.3342.61142.435.9316.49495.334.27172.0
DepthAnything3Rel (%)↓\downarrow116.118.694.3286.7517.1322.477146.57.54938.69
DepthAnything3δ1.05\delta_{1.05} (%)↑\uparrow17.0444.4091.5683.9092.1895.660.291688.4164.18
DepthAnything3δ1.10\delta_{1.10} (%)↑\uparrow38.1457.9694.3290.7095.1397.042.99093.2671.19
DepthAnything3δ1.25\delta_{1.25} (%)↑\uparrow47.6667.9795.9496.8296.1497.9612.6695.2176.29
MapAnythingMAE (cm)↓\downarrow37.4515.3114.7834.3914.6619.1754.4718.5326.09
MapAnythingRMSE (cm)↓\downarrow46.0519.8230.3745.8028.0926.6569.7437.1637.96
MapAnythingRel (%)↓\downarrow15.107.5805.14510.325.8606.19321.499.50910.15
MapAnythingδ1.05\delta_{1.05} (%)↑\uparrow8.76137.0581.3432.7469.7952.518.28967.9644.80
MapAnythingδ1.10\delta_{1.10} (%)↑\uparrow34.2771.4490.3864.5885.9588.7323.7487.8968.37
MapAnythingδ1.25\delta_{1.25} (%)↑\uparrow73.2396.1294.8092.3697.2896.0954.1791.3786.92
LongLRMMAE (cm)↓\downarrow14.2511.0110.0417.7011.559.82519.1611.1513.09
LongLRMRMSE (cm)↓\downarrow18.2014.3016.2325.5320.2014.5325.5817.4819.00
LongLRMRel (%)↓\downarrow6.2815.9103.4286.5284.5303.2728.8297.0225.724
LongLRMδ1.05\delta_{1.05} (%)↑\uparrow46.1855.5682.1762.6080.2785.0944.2470.8965.87
LongLRMδ1.10\delta_{1.10} (%)↑\uparrow80.1680.6895.1387.4891.8596.2467.9884.6585.52
LongLRMδ1.25\delta_{1.25} (%)↑\uparrow98.8598.4299.1297.5997.8999.0492.8193.4697.14
VoxelTTO (w/o TTO)MAE (cm)↓\downarrow2.9471.4404.2973.1281.7741.6152.7822.9322.615
VoxelTTO (w/o TTO)RMSE (cm)↓\downarrow7.3953.84915.0610.0811.395.5727.08311.839.032
VoxelTTO (w/o TTO)Rel (%)↓\downarrow1.2990.75621.4111.4750.65570.54551.1731.6111.115
VoxelTTO (w/o TTO)δ1.05\delta_{1.05} (%)↑\uparrow97.0098.5497.4998.2899.3899.2997.6697.5298.14
VoxelTTO (w/o TTO)δ1.10\delta_{1.10} (%)↑\uparrow98.1099.1098.2998.6199.4499.5099.3098.3198.83
VoxelTTO (w/o TTO)δ1.25\delta_{1.25} (%)↑\uparrow99.3899.8398.8499.1299.5699.6999.7498.9399.38
OursMAE (cm)↓\downarrow1.9401.2793.4222.2812.6861.1921.9591.3562.014
OursRMSE (cm)↓\downarrow5.7664.08211.698.0058.3604.8474.2758.5366.945
OursRel (%)↓\downarrow0.84730.68231.1431.1560.96320.39540.85870.88490.866
Oursδ1.05\delta_{1.05} (%)↑\uparrow98.2498.4498.4398.7799.6199.4799.2998.7198.86
Oursδ1.10\delta_{1.10} (%)↑\uparrow98.9099.0198.9299.0199.6899.6499.7399.0799.24
Oursδ1.25\delta_{1.25} (%)↑\uparrow99.6399.7699.2899.3099.7699.7499.9199.3199.58

Table A4. Per-scene depth view synthesis results on Replica test views.

For pose-free baselines, we additionally align the rendered depth scale to the ground-truth depth before computing depth metrics. Inaccurate predicted camera poses may make the scale estimated through camera-based alignment unreliable. As shown in Equation (15), for each rendered depth map from pose-free baselines, we estimate a scalar a∗a^{*} by minimizing the squared error over V\mathcal{V}, which contains the pixels with valid depth values in the ground-truth depth map.

a∗=arg⁡min⁡a>0∑u∈V(ad^u−du)2=∑u∈Vd^udu∑u∈Vd^u2.(15)a^{*} = \underset{a>0}{\arg\min}\sum_{u\in\mathcal{V}}\left(a\widehat{d}_u-d_u\right)^2 = \frac{\sum_{u\in\mathcal{V}}\widehat{d}_u d_u}{\sum_{u\in\mathcal{V}}\widehat{d}_u^2}. \tag*{(15)}
Additional qualitative results for novel view synthesis

Figure A1. Additional qualitative results for novel view synthesis.

Here, d^u\hat{d}_u and dud_u are the rendered and ground-truth depths at pixel uu. We compute depth metrics using d^ualigned=a∗d^u\hat{d}_u^{\mathrm{aligned}} = a^* \hat{d}_u on valid pixels V\mathcal{V}. For evaluation on Tanks and Temples (TAT), we use the

Additional qualitative results for depth view synthesis

Figure A2. Additional qualitative results for depth view synthesis.

ScanVoxelTTO (w/o TTO)VoxelTTO (Camera Enc.)VoxelTTO
Acc.↓Com.↓Over.↓Acc.↓Com.↓Over.↓Acc.↓Com.↓Over.↓
scan244.7414.6014.6711.5522.4341.9933.4834.9454.214
scan3711.17510.74510.9609.22410.2009.7125.0866.6365.861
scan403.2275.4564.3411.9983.2262.6120.7212.1411.431
scan556.0459.9537.9992.0632.2912.1771.3952.4251.910
scan635.3567.2366.2964.6876.0795.3832.3932.4182.405
scan654.4579.7137.0852.4513.3832.9172.0442.2672.155
scan694.5264.2734.4005.3585.6045.4812.2732.8662.569
scan839.50413.36311.4335.33011.2558.2934.6776.8995.788
scan975.56111.4838.5225.1088.4246.7664.9586.6785.818
scan1057.92915.64811.7893.7534.6484.2001.3742.0521.713
scan1064.35710.6197.4882.9684.1253.5473.7024.9814.342
scan1103.2923.8653.5792.2021.8972.0502.4183.0352.726
scan1144.86114.5019.6812.1962.3252.2603.9514.2914.121
scan1184.0228.6436.3323.4803.4903.4852.7903.3613.075
scan1224.2827.2835.7823.4163.4073.4113.0103.2053.107
Average5.5569.1597.3573.7194.8534.2862.9523.8803.416

Table A11. Per-scan Chamfer distances (mm) for surface alignment on DTU. Lower is better; the best result for each metric and scan is bold.

3 scenes included in RobustMVD [39]. We likewise sample input views at uniform intervals with a stride of 3 frames and use all remaining frames within the sampled range as test views. Detailed RGB view synthesis results on TAT dataset are shown in Table A5 and Table A6. Detailed depth view synthesis results on TAT dataset are shown in Table A7 and Table A8.

Table A5. Per-scene RGB view synthesis results on TAT input views.

MethodMetricBarnChurchCourthouseAverage
VolSplatPSNR↑\mathrm{PSNR}\uparrow14.1516.0916.1215.45
SSIM↑\mathrm{SSIM}\uparrow0.5430.5380.7040.595
LPIPS↓\mathrm{LPIPS}\downarrow0.4310.4740.3510.419
MVSplatPSNR↑\mathrm{PSNR}\uparrow15.3716.3417.7316.48
SSIM↑\mathrm{SSIM}\uparrow0.5940.5340.7160.615
LPIPS↓\mathrm{LPIPS}\downarrow0.4100.4240.3060.380
AnySplatPSNR↑\mathrm{PSNR}\uparrow15.8118.5317.6717.34
SSIM↑\mathrm{SSIM}\uparrow0.6460.5800.7080.645
LPIPS↓\mathrm{LPIPS}\downarrow0.2960.3180.3200.311
YoNoSplatPSNR↑\mathrm{PSNR}\uparrow12.6214.1014.1113.61
SSIM↑\mathrm{SSIM}\uparrow0.4750.3460.6090.477
LPIPS↓\mathrm{LPIPS}\downarrow0.5010.5900.4570.516
DepthAnything3PSNR↑\mathrm{PSNR}\uparrow19.6117.4620.5319.20
SSIM↑\mathrm{SSIM}\uparrow0.6280.5030.7320.621
LPIPS↓\mathrm{LPIPS}\downarrow0.3320.4050.3070.348
MapAnythingPSNR↑\mathrm{PSNR}\uparrow9.9814.4113.5912.66
SSIM↑\mathrm{SSIM}\uparrow0.4600.4110.6210.497
LPIPS↓\mathrm{LPIPS}\downarrow0.5930.6320.4860.570
LongLRMPSNR↑\mathrm{PSNR}\uparrow21.8220.6623.5522.01
SSIM↑\mathrm{SSIM}\uparrow0.6950.6110.8090.705
LPIPS↓\mathrm{LPIPS}\downarrow0.2790.3720.2540.302
VoxelTTO (w/o TTO)PSNR↑\mathrm{PSNR}\uparrow22.3022.0326.9023.74
SSIM↑\mathrm{SSIM}\uparrow0.7250.7070.8690.767
LPIPS↓\mathrm{LPIPS}\downarrow0.2780.2740.1890.247
OursPSNR↑\mathrm{PSNR}\uparrow20.3818.0124.3020.89
SSIM↑\mathrm{SSIM}\uparrow0.6340.5200.8130.656
LPIPS↓\mathrm{LPIPS}\downarrow0.3550.4320.2540.347

Table A5. Per-scene RGB view synthesis results on TAT input views.

Table A6. Per-scene RGB view synthesis results on TAT test views.

MethodMetricBarnChurchCourthouseAverage
VolSplatPSNR↑\mathrm{PSNR}\uparrow15.9215.9717.1016.33
SSIM↑\mathrm{SSIM}\uparrow0.7860.5730.8400.733
LPIPS↓\mathrm{LPIPS}\downarrow0.1810.3650.1770.241
MVSplatPSNR↑\mathrm{PSNR}\uparrow13.6016.0915.4715.05
SSIM↑\mathrm{SSIM}\uparrow0.4950.4230.6490.522
LPIPS↓\mathrm{LPIPS}\downarrow0.4770.5040.4000.460
AnySplatPSNR↑\mathrm{PSNR}\uparrow11.6714.6311.8112.71
SSIM↑\mathrm{SSIM}\uparrow0.4050.3650.5500.440
LPIPS↓\mathrm{LPIPS}\downarrow0.4400.5260.4690.478
YoNoSplatPSNR↑\mathrm{PSNR}\uparrow12.1712.3015.4113.29
SSIM↑\mathrm{SSIM}\uparrow0.6680.4910.8220.660
LPIPS↓\mathrm{LPIPS}\downarrow0.3000.4560.1550.304
DepthAnything3PSNR↑\mathrm{PSNR}\uparrow17.2216.5919.2317.68
SSIM↑\mathrm{SSIM}\uparrow0.5530.4110.7040.556
LPIPS↓\mathrm{LPIPS}\downarrow0.4040.5090.3250.413
MapAnythingPSNR↑\mathrm{PSNR}\uparrow13.6815.3414.9714.66
SSIM↑\mathrm{SSIM}\uparrow0.5510.5680.6580.592
LPIPS↓\mathrm{LPIPS}\downarrow0.4470.3700.3270.381
LongLRMPSNR↑\mathrm{PSNR}\uparrow10.4615.0311.0512.18
SSIM↑\mathrm{SSIM}\uparrow0.3830.3300.5300.415
LPIPS↓\mathrm{LPIPS}\downarrow0.5880.6140.5990.600
VoxelTTO (w/o TTO)PSNR↑\mathrm{PSNR}\uparrow13.6714.9315.5314.71
SSIM↑\mathrm{SSIM}\uparrow0.4530.3410.6610.485
LPIPS↓\mathrm{LPIPS}\downarrow0.4770.5600.3880.475
OursPSNR↑\mathrm{PSNR}\uparrow19.3116.1622.2119.23
SSIM↑\mathrm{SSIM}\uparrow0.5610.3690.7650.565
LPIPS↓\mathrm{LPIPS}\downarrow0.3780.5100.2860.391

Table A6. Per-scene RGB view synthesis results on TAT test views.

MethodMetricBarnChurchCourthouseAverage
VolSplatMAE (cm)↓\downarrow289.1350.8472.0370.7
VolSplatRMSE (cm)↓\downarrow378.2423.5670.4490.7
VolSplatRel (%)↓\downarrow31.8439.5929.7933.73
VolSplatδ1.05\delta_{1.05} (%)↑\uparrow11.067.64311.259.984
VolSplatδ1.10\delta_{1.10} (%)↑\uparrow21.3314.7521.3419.13
VolSplatδ1.25\delta_{1.25} (%)↑\uparrow45.6731.8946.2041.25
MVSplatMAE (cm)↓\downarrow536.6619.21214789.9
MVSplatRMSE (cm)↓\downarrow661.7663.41358894.3
MVSplatRel (%)↓\downarrow61.4072.1282.4271.98
MVSplatδ1.05\delta_{1.05} (%)↑\uparrow0.90331.5750.25680.911
MVSplatδ1.10\delta_{1.10} (%)↑\uparrow1.7913.1310.50581.809
MVSplatδ1.25\delta_{1.25} (%)↑\uparrow5.0107.9911.2184.739
AnySplatMAE (cm)↓\downarrow42.7232.2091.5955.50
AnySplatRMSE (cm)↓\downarrow66.9050.98234.5117.4
AnySplatRel (%)↓\downarrow4.8224.1075.4504.793
AnySplatδ1.05\delta_{1.05} (%)↑\uparrow61.7573.7374.5770.01
AnySplatδ1.10\delta_{1.10} (%)↑\uparrow93.3890.3889.1990.98
AnySplatδ1.25\delta_{1.25} (%)↑\uparrow97.9898.6994.2196.95
YoNoSplatMAE (cm)↓\downarrow306.1135.6856.8432.8
YoNoSplatRMSE (cm)↓\downarrow369.3183.6981.8511.5
YoNoSplatRel (%)↓\downarrow34.5216.3360.3437.06
YoNoSplatδ1.05\delta_{1.05} (%)↑\uparrow3.96319.120.97488.018
YoNoSplatδ1.10\delta_{1.10} (%)↑\uparrow7.11437.942.03715.69
YoNoSplatδ1.25\delta_{1.25} (%)↑\uparrow17.2678.135.24633.54
DepthAnything3MAE (cm)↓\downarrow110.0830.5173.5371.3
DepthAnything3RMSE (cm)↓\downarrow183.9853.5333.4456.9
DepthAnything3Rel (%)↓\downarrow12.61100.0011.7441.44
DepthAnything3δ1.05\delta_{1.05} (%)↑\uparrow22.18034.1118.76
DepthAnything3δ1.10\delta_{1.10} (%)↑\uparrow43.97066.1836.71
DepthAnything3δ1.25\delta_{1.25} (%)↑\uparrow92.41088.6260.34
MapAnythingMAE (cm)↓\downarrow102.220.2789.8770.77
MapAnythingRMSE (cm)↓\downarrow161.430.48227.8139.9
MapAnythingRel (%)↓\downarrow11.362.5025.8216.560
MapAnythingδ1.05\delta_{1.05} (%)↑\uparrow29.1690.9080.9867.01
MapAnythingδ1.10\delta_{1.10} (%)↑\uparrow60.0197.1586.1981.11
MapAnythingδ1.25\delta_{1.25} (%)↑\uparrow92.0299.5892.7394.77
LongLRMMAE (cm)↓\downarrow150.4160.8198.3169.8
LongLRMRMSE (cm)↓\downarrow157.2187.5200.2181.6
LongLRMRel (%)↓\downarrow21.8025.8226.4124.67
LongLRMδ1.05\delta_{1.05} (%)↑\uparrow0.708012.071.4964.757
LongLRMδ1.10\delta_{1.10} (%)↑\uparrow3.74623.832.2449.940
LongLRMδ1.25\delta_{1.25} (%)↑\uparrow69.7553.1940.3654.43
VoxelTTO (w/o TTO)MAE (cm)↓\downarrow17.808.68948.0024.83
VoxelTTO (w/o TTO)RMSE (cm)↓\downarrow66.7621.58157.982.07
VoxelTTO (w/o TTO)Rel (%)↓\downarrow2.0511.1743.0122.079
VoxelTTO (w/o TTO)δ1.05\delta_{1.05} (%)↑\uparrow97.4996.7694.8196.35
VoxelTTO (w/o TTO)δ1.10\delta_{1.10} (%)↑\uparrow98.3298.5595.4897.45
VoxelTTO (w/o TTO)δ1.25\delta_{1.25} (%)↑\uparrow99.4299.5896.4798.49
OursMAE (cm)↓\downarrow21.7220.1353.6831.84
OursRMSE (cm)↓\downarrow77.5636.63157.790.63
OursRel (%)↓\downarrow2.3182.6603.6492.875
Oursδ1.05\delta_{1.05} (%)↑\uparrow94.4386.5089.4590.12
Oursδ1.10\delta_{1.10} (%)↑\uparrow97.6694.8994.9595.83
Oursδ1.25\delta_{1.25} (%)↑\uparrow99.2498.8096.4398.15

Table A7. Per-scene depth view synthesis results on TAT input views.

MethodMetricBarnChurchCourthouseAverage
VolSplatMAE (cm)↓\downarrow34.4855.1866.2751.98
VolSplatRMSE (cm)↓\downarrow41.8768.7596.8469.15
VolSplatRel (%)↓\downarrow25.5841.0826.5131.05
VolSplatδ1.05\delta_{1.05} (%)↑\uparrow10.346.36510.479.058
VolSplatδ1.10\delta_{1.10} (%)↑\uparrow20.3811.9721.6918.01
VolSplatδ1.25\delta_{1.25} (%)↑\uparrow47.3226.1554.6342.70
MVSplatMAE (cm)↓\downarrow100.3105.6171.3125.7
MVSplatRMSE (cm)↓\downarrow117.2111.3202.7143.7
MVSplatRel (%)↓\downarrow74.2681.1274.3576.57
MVSplatδ1.05\delta_{1.05} (%)↑\uparrow0.050920.53510.44540.343
MVSplatδ1.10\delta_{1.10} (%)↑\uparrow0.14201.0770.87830.699
MVSplatδ1.25\delta_{1.25} (%)↑\uparrow0.62063.8902.0992.203
AnySplatMAE (cm)↓\downarrow638.6737.531.71469.3
AnySplatRMSE (cm)↓\downarrow685.4754.757.45499.1
AnySplatRel (%)↓\downarrow484.4584.212.42360.3
AnySplatδ1.05\delta_{1.05} (%)↑\uparrow0016.225.405
AnySplatδ1.10\delta_{1.10} (%)↑\uparrow0046.1215.37
AnySplatδ1.25\delta_{1.25} (%)↑\uparrow0091.2330.40
YoNoSplatMAE (cm)↓\downarrow52.3394.5266.8871.24
YoNoSplatRMSE (cm)↓\downarrow62.11100.588.0983.57
YoNoSplatRel (%)↓\downarrow40.0373.6229.1347.59
YoNoSplatδ1.05\delta_{1.05} (%)↑\uparrow8.9382.01311.497.480
YoNoSplatδ1.10\delta_{1.10} (%)↑\uparrow19.144.17722.8315.38
YoNoSplatδ1.25\delta_{1.25} (%)↑\uparrow48.2410.3857.5138.70
DepthAnything3MAE (cm)↓\downarrow351.9160.170.36194.1
DepthAnything3RMSE (cm)↓\downarrow512.4271.5114.5299.4
DepthAnything3Rel (%)↓\downarrow39.9320.0931.4830.49
DepthAnything3δ1.05\delta_{1.05} (%)↑\uparrow0.863010.680.26423.934
DepthAnything3δ1.10\delta_{1.10} (%)↑\uparrow1.58122.380.58358.181
DepthAnything3δ1.25\delta_{1.25} (%)↑\uparrow3.09664.931.85223.29
MapAnythingMAE (cm)↓\downarrow716.8744.521.43494.2
MapAnythingRMSE (cm)↓\downarrow729.9757.749.87512.4
MapAnythingRel (%)↓\downarrow548.1641.18.352399.1
MapAnythingδ1.05\delta_{1.05} (%)↑\uparrow0045.8615.28
MapAnythingδ1.10\delta_{1.10} (%)↑\uparrow0084.1928.06
MapAnythingδ1.25\delta_{1.25} (%)↑\uparrow0093.9531.31
LongLRMMAE (cm)↓\downarrow81.74176.9266.0174.9
LongLRMRMSE (cm)↓\downarrow284.7224.5877.0462.1
LongLRMRel (%)↓\downarrow9.16523.0716.8016.34
LongLRMδ1.05\delta_{1.05} (%)↑\uparrow57.8618.2750.1442.08
LongLRMδ1.10\delta_{1.10} (%)↑\uparrow78.9032.7466.4759.37
LongLRMδ1.25\delta_{1.25} (%)↑\uparrow93.5259.9188.1780.53
VoxelTTO (w/o TTO)MAE (cm)↓\downarrow58.2446.6419.2041.36
VoxelTTO (w/o TTO)RMSE (cm)↓\downarrow86.9465.8039.0663.93
VoxelTTO (w/o TTO)Rel (%)↓\downarrow6.6076.7447.7567.035
VoxelTTO (w/o TTO)δ1.05\delta_{1.05} (%)↑\uparrow52.7967.3937.0152.39
VoxelTTO (w/o TTO)δ1.10\delta_{1.10} (%)↑\uparrow76.6880.8179.0378.84
VoxelTTO (w/o TTO)δ1.25\delta_{1.25} (%)↑\uparrow94.2296.4095.7195.44
OursMAE (cm)↓\downarrow24.9625.927.91919.60
OursRMSE (cm)↓\downarrow69.7545.9123.0146.22
OursRel (%)↓\downarrow2.7823.4383.4883.235
Oursδ1.05\delta_{1.05} (%)↑\uparrow90.9984.0485.2486.75
Oursδ1.10\delta_{1.10} (%)↑\uparrow97.1191.7496.6395.16
Oursδ1.25\delta_{1.25} (%)↑\uparrow99.4098.1497.4498.33

Table A8. Per-scene depth view synthesis results on TAT test views.

For evaluation on RE10K[61], we use 10 scenes included in FineSplat [23]. We likewise sample input views at uniform intervals with a stride of 10 frames and use all remaining frames within the sampled range as test views. Detailed RGB view synthesis results on RE10K dataset are shown in Table A9 and Table A10. Since RE10K is an RGB dataset, we did not evaluate the accuracy of depth view synthesis.

MethodMetric89999a83cb78ec832a7918811fcee38d0cc7cbbf749f321b42fb0ecbb90b641a12a0a7760e3fc05aAverage
VolSplatPSNR↑16.6822.6121.7521.8816.6119.1619.7718.8024.0524.1620.55
SSIM↑0.5940.8190.8040.8170.5830.6390.7410.6620.8060.8410.731
LPIPS↓0.4580.3940.3690.3810.5330.5210.4230.4260.2930.3500.415
MVSplatPSNR↑11.2414.5316.7312.1414.9311.4118.4612.5215.5219.7714.72
SSIM↑0.4940.7230.7170.7270.5710.5510.7440.4880.6260.8390.648
LPIPS↓0.5480.4270.3760.4160.4990.5300.4180.4910.4090.2960.441
AnySplatPSNR↑19.5021.0319.6432.6521.8525.1419.9424.2327.1724.9523.61
SSIM↑0.7600.8520.7660.9620.7970.8720.7680.8590.8780.9040.842
LPIPS↓0.2230.2460.3070.1310.2310.1880.3280.1380.1520.2190.216
YoNoSplatPSNR↑22.0630.2525.7532.0322.5120.0026.8422.4622.2929.2825.35
SSIM↑0.8080.9470.8850.9440.7760.6950.8870.7810.7350.9090.837
LPIPS↓0.1840.1460.1730.1390.2170.3110.1660.1650.2180.1680.189
DepthAnything3PSNR↑19.1124.6520.3625.9417.5624.0624.3422.7121.8924.0222.47
SSIM↑0.7800.9200.8010.9180.6460.8540.8590.8150.7150.8650.818
LPIPS↓0.2170.1850.2770.1830.3680.1990.1930.1650.2810.2630.233
MapAnythingPSNR↑12.2611.9615.8611.6315.2913.3715.7314.5814.4612.4613.76
SSIM↑0.4920.5720.6550.5380.6020.5340.6650.5700.6330.5710.583
LPIPS↓0.5810.6230.5620.6270.5300.6450.5630.5440.5420.6700.589
LongLRMPSNR↑24.5929.5325.5329.7523.1526.5423.8024.3725.4028.5426.12
SSIM↑0.8830.9490.9080.9440.8470.9070.8390.8350.8820.9330.893
LPIPS↓0.1710.1820.2120.1870.2420.1890.2590.1830.2080.2250.206
VoxelTTO (w/o TTO)PSNR↑22.1627.9923.7231.0525.9226.2826.4525.6928.9131.5126.97
SSIM↑0.8320.9290.8510.9520.8710.8810.8860.8770.9020.9490.893
LPIPS↓0.2290.2320.2800.1640.2250.2390.2360.1940.1800.1770.216
OursPSNR↑22.2828.1225.7430.3426.1226.2826.2526.1029.3830.7627.14
SSIM↑0.8440.9390.9050.9430.8760.8820.8840.8860.9130.9420.901
LPIPS↓0.2250.2220.2240.1710.2140.2440.2360.1890.1630.1830.207

Table A9. Per-scene RGB view synthesis results on RE10K input views.

MethodMetric89999a83cb78ec832a7918811fcee38d0cc7cbbf749f321b42fb0ecbb90b641a12a0a7760e3fc05aAverage
VolSplatPSNR↑11.4114.8813.6414.4312.3512.7717.2114.1014.5515.9314.13
SSIM↑0.4230.6370.5880.6080.4070.4590.6240.5060.5830.6830.552
LPIPS↓0.5700.5350.5200.5320.6360.6200.5230.5520.4930.4810.546
MVSplatPSNR↑11.3314.7216.6912.5814.7311.3516.5012.7615.4120.3114.64
SSIM↑0.4660.6920.6870.7190.5230.4940.6670.4420.5890.8300.611
LPIPS↓0.5770.4720.4200.4210.5430.5790.5480.5280.4500.3400.488
AnySplatPSNR↑10.589.5413.6312.3511.519.4714.4812.4315.2812.2012.15
SSIM↑0.3850.5750.5670.6760.4350.4290.5710.4690.5620.6630.533
LPIPS↓0.6070.5160.5290.4710.6220.5910.5460.5090.4200.5280.534
YoNoSplatPSNR↑8.428.907.8510.218.327.989.146.076.909.608.34
SSIM↑0.5080.7140.5470.7810.5060.6000.6110.4640.4990.7740.601
LPIPS↓0.7020.5900.6470.6020.6950.6730.6830.7140.7010.6240.663
DepthAnything3PSNR↑15.9820.8919.7914.4117.0317.6620.0416.9019.6816.5417.89
SSIM↑0.5830.8260.7730.7280.6190.6350.7380.6040.6500.7850.694
LPIPS↓0.3170.2480.3000.5110.3920.3920.3020.3020.3420.4450.355
MapAnythingPSNR↑11.379.9812.169.0510.409.1912.3910.2111.104.2310.01
SSIM↑0.4090.5260.5110.5150.3780.3740.5350.3620.4030.2900.430
LPIPS↓0.6250.6870.6400.6910.6970.7060.6460.6740.6610.7060.673
LongLRMPSNR↑15.2414.9016.8619.7218.9118.4720.1320.3515.2914.5917.44
SSIM↑0.7670.8410.8250.8880.7940.8130.7900.7980.8100.8150.814
LPIPS↓0.3130.3770.3060.3080.3020.3250.3210.2410.2650.4600.322
VoxelTTO (w/o TTO)PSNR↑13.5119.2416.6115.2819.4516.6019.5017.7517.4119.2817.46
SSIM↑0.4950.7780.6440.7370.6880.6170.7120.6300.6340.7990.673
LPIPS↓0.4400.3240.4450.4610.3080.4420.3520.3050.3570.3470.378
OursPSNR↑17.6422.8919.5419.9021.1820.1322.6821.1122.4621.8120.93
SSIM↑0.6790.8600.7300.8010.7350.6950.7930.7620.7650.8240.764
LPIPS↓0.2810.2660.3120.3450.2600.3220.2890.2250.2130.2770.279

Table A10. Per-scene RGB view synthesis results on RE10K test views.

Per-scan surface-alignment results on DTU are shown in Figure A11.

References

  1. [1]Henrik Aanaes, Rasmus R. Jensen, George Vogiatzis, Engin Tola, and Anders B. Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153–168, 2016. 5, 7, 8DOI
  2. [2]Byeongjoo Ahn, Michael De Zeeuw, Ioannis Gkioulekas, and Aswin C. Sankaranarayanan. Neural kaleidoscopic space sculpting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4349–4358, 2023. 1DOI
  3. [3]Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2arxiv.org/abs/2111.12077
  4. [4]Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitscenes–a diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 5, 8arxiv.org/abs/2111.08897
  5. [5]David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 2, 5arxiv.org/abs/2312.12337
  6. [6]Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2, 5DOI
  7. [7]Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022. 3
  8. [8]Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2arxiv.org/abs/2404.17774
  9. [9]Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224–236, 2018. 1arxiv.org/abs/1712.07629
  10. [10]Sven Elflein, Ruilong Li, Sérgio Agostinho, Zan Gojcic, Laura Leal-Taixé, Qunije Zhou, and Aljosa Osep. Vgg-t³: Offline feed-forward 3d reconstruction at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 3
  11. [11]Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796–20805, 2024. 2arxiv.org/abs/2312.07504
  12. [12]Diego Gomez, Antoine Guédon, Nissim Maruani, Bingchen Gong, and Maks Ovsjanikov. From blobs to spokes: High-fidelity surface reconstruction via oriented gaussians, 2026. 2arxiv.org/abs/2604.07337
  13. [13]Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 2
  14. [14]Seongbo Ha, Jiung Yeon, and Hyunwoo Yu. Rgbd gs-icp slam. In European conference on computer vision, pages 180–197. Springer, 2024. 2
  15. [15]Dongchen Han, Yingxin Li, Tianyu Li, Zixuan Cao, Ziming Wang, Jun Song, Yu Cheng, Bo Zheng, and Gao Huang. Vit³: Unlocking test-time training in vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026. 3DOI
  16. [16]Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4arxiv.org/abs/2106.09685
  17. [17]Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2DOI
  18. [18]Han Huang, Yulun Wu, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. Fatsgs: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2
  19. [19]Lihan Jiang, Yucheng Mao, Linning Xu, Tao Lu, Kerui Ren, Yichen Jin, Xudong Xu, Mulin Yu, Jiangmiao Pang, Feng Zhao, et al. Anysplat: Feed-forward 3d gaussian splatting from unconstrained views. ACM Transactions on Graphics (TOG), 44(6):1–16, 2025. 2, 3, 5, 8
  20. [20]Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21357–21366, 2024. 2arxiv.org/abs/2312.02126
  21. [21]Nikhil Keetha, Norman Müller, Johannes Schönberger, Lorenzo Porzi, Yuchen Zhang, Tobias Fischer, Arno Knापitsch, Duncan Zauss, Ethan Weber, Nelson Antunes, Jonathon Luiten, Manuel Lopez-Antequera, Samuel Rota Bulò, Christian Richardt, Deva Ramanan, Sebastian Scherer, and Peter Kontschieder. MapAnything: Universal feed-forward metric 3D reconstruction. In International Conference on 3D Vision (3DV). IEEE, 2026. 2, 3, 5arxiv.org/abs/2509.13414
  22. [22]Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 5arxiv.org/abs/2308.04079
  23. [23]Hyejin Kim, Hwasik Jeong, Joohyuk Park, Gyeongjin Kang, Jungwoo Kim, Younghun Kim, and Eunbyung Park. Finesplat: Fine patches, local groups, detailed splats, 2026. 12arxiv.org/abs/2610.25011
  24. [24]Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5, 8
  25. [25]Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2arxiv.org/abs/2406.09756
  26. [26]Yinglong Li, Donghui Shen, Xiaoyu Zhang, Zhichao Ye, Hongyu Wu, Aimin Hao, Guofeng Zhang, and Haomin Liu. Querysplat: Decoupling geometry and appearance representations in 3dgs prediction. arXiv preprint arXiv:2608.01186, 2026. 3arxiv.org/abs/2608.01186
  27. [27]Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647, 2025. 2, 3, 4, 5, 7, 1arxiv.org/abs/2511.10647
  28. [28]Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyiu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22160–22169. IEEE, 2024. 5, 8arxiv.org/abs/2312.16256
  29. [29]Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 2arxiv.org/abs/2312.00109
  30. [30]Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2arxiv.org/abs/2003.08934
  31. [31]Bailey Miller, Hanyu Chen, Alice Lai, and Ioannis Gkioulekas. Objects as volumes: A stochastic geometry view of opaque solids. IEEE, 2023. 2, 3arxiv.org/abs/2312.15406
  32. [32]Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1–102:15, 2022. 2DOI
  33. [33]Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bo-janowski. DinoV2: Learning robust visual features without supervision, 2023. 3
  34. [34]Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, Alejandro Newell, Hei Law, Ankit Goyal, Kaiyu Yang, and Jia Deng. Infinite photorealistic worlds using procedural generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12630–12641, 2023. 5, 8arxiv.org/abs/2306.09310
  35. [35]Jiawei Ren, Michal Tyszkiewicz, Jiahui Huang, and Zan Gojcic. Tokens: Decoupling 3d gaussian prediction from pixels with learnable tokens. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026. 3arxiv.org/abs/2604.15239
  36. [36]Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1arxiv.org/abs/1911.11763
  37. [37]Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2DOI
  38. [38]Johannes Lutz Schönberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 2DOI
  39. [39]Philipp Schröppel, Jan Bechtold, Artemij Amiranshili, and Thomas Brox. A benchmark and a baseline for robust multi-view depth estimation. In Proceedings of the International Conference on 3D Vision (3DV), 2022. 12arxiv.org/abs/2209.06681
  40. [40]Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqiang Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strastad, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5, 8, 2arxiv.org/abs/1906.05797
  41. [41]Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on pattern analysis and machine intelligence, 13(4): 376–380, 1991. 3, 2DOI
  42. [42]Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2arxiv.org/abs/2408.16061
  43. [43]Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsf: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686–21697, 2024. 2arxiv.org/abs/2312.04563
  44. [44]Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1, 3, 5arxiv.org/abs/2503.11651
  45. [45]Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bo- janowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. Vggt-ω, 2026. 2
  46. [46]Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2arxiv.org/abs/2106.10689
  47. [47]Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 3arxiv.org/abs/2312.14132
  48. [48]Weijie Wang, Yeqing Chen, Zeyu Zhang, Hengyu Liu, Haoxiao Wang, Zhiyuan Feng, Wenkang Qin, Feng Chen, Zheng Zhu, Donny Y. Chen, and Bohan Zhuang. Volsplat: Rethinking feed-forward 3d gaussian splatting with voxel-aligned prediction. arXiv preprint arXiv:2509.19297, 2025. 3, 5arxiv.org/abs/2509.19297
  49. [49]Yiming Wang, Qin Han, Marc Habermann, Kostas Daniilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2arxiv.org/abs/2212.05231
  50. [50]Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, Chunhua Shen, and Tong He. π³: Permutation-Equivariant Visual Geometry Learning. arXiv preprint arXiv:2507.13347, 2025. 2arxiv.org/abs/2507.13347
  51. [51]Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 5pubmed.ncbi.nlm.nih.gov/15376593
  52. [52]Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729–21740, 2023. 1arxiv.org/abs/2303.09551
  53. [53]Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. In CVPR, 2025. 2arxiv.org/abs/2410.13862
  54. [54]Jianing Yang, Alexander Sax, Kevin J. Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2arxiv.org/abs/2501.13928
  55. [55]Sizhe Yang, Linning Xu, Hao Li, Juncheng Mu, Jia Zeng, Dahua Lin, and Jiangmiao Pang. Robo3r: Enhancing robotic manipulation with accurate feed-forward 3d reconstruction. In Proceedings of Robotics: Science and Systems, Sydney, Australia, 2026. 1arxiv.org/abs/2602.10101
  56. [56]Botao Ye, Boqi Chen, Haofei Xu, Daniel Barath, and Marc Pollefeys. Yonosplat: You only need one model for feedforward 3d gaussian splatting, 2025. 2, 5arxiv.org/abs/2511.07321
  57. [57]Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12–22. IEEE, 2023. 5, 8arxiv.org/abs/2308.11417
  58. [58]Baowen Zhang, Chenxing Jiang, Heng Li, Shaojie Shen, and Ping Tan. Geometry-grounded gaussian splatting. arXiv preprint arXiv:2601.17835, 2026. 2, 3, 4arxiv.org/abs/2601.17835
  59. [59]Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5arxiv.org/abs/1801.03924
  60. [60]Lei Zhou, Haoyu Wu, Akshat Dave, and Dimitris Samaras. Learning 3d reconstruction with priors in test time. arXiv preprint arXiv:2604.03878, 2026. 3arxiv.org/abs/2604.03878
  61. [61]Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 5, 8, 12arxiv.org/abs/1805.09817
  62. [62]Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4349–4359. IEEE, 2025. 5arxiv.org/abs/2410.12781

Paper details

vlnc:W_e832nq2r
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

ChatGPT

View statement
Looking for related papers…
Contents