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.

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 can still differ substantially from . 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 to learn cross-view visual representations from uncalibrated images . Separate prediction heads recover 3D scene information, including point maps , depth maps , and camera parameters , as shown in Equation (1).
3DGS VFMs typically upsample patch tokens with DPT to obtain pixel-aligned Gaussian attributes (center , opacity , rotation quaternion , scale , and SH coefficients ). 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 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 can still differ substantially from . 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 , camera parameters , and confidence map for each view. We add a feature DPT that upsamples patch tokens into dense features . We then back-project pixels into world coordinates, obtaining a dense feature point cloud , where is the number of points and , , , and denote point coordinate, color, feature, and confidence. To reduce memory use and avoid excessive overlap, we aggregate the point cloud into voxel features. Let denote the points within voxel . Its feature center is , which is not constrained to the regular voxel center and thus preserves geometric accuracy. Voxel color and feature are also confidence-weighted averages. We define voxel confidence as 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 into 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, and . Centers, rotation quaternions, scales, and opacities of Gaussian primitives are decoded by . Second-order spherical-harmonic coefficients are decoded by .
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 , volume rendering of the stochastic solid is equivalent to 3DGS rasterization [58], as shown in Equation (2), where is the ray direction, is the stochastic solid attenuation coefficient at point along the ray, is accumulated transmittance at position , and is the directional color.

Figure 2. Overview of the VoxelTTO pipeline.
An appropriate vacancy is required to obtain images equivalent to 3DGS rasterization. GGGS[58] proves that the renderings are equivalent when . The accumulated transmittance is then computed as shown in Equation (3), where is the location of the Gaussian maximum along the ray.
For depth rendering, we use the median depth, defined by the point where with a fixed number of bisection iterations. During backpropagation, we differentiate the fixed condition with respect to Gaussian parameters , as shown in Equation (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.

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.

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 and ground-truth rotations , as shown in Equation (5).
To address scale ambiguity, we normalize each translation set by mean scale , 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 , and we directly apply SmoothL1 to the predicted and ground-truth values. The camera-based TTO objective is .
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 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 , cosine annealing, 2K warm-up iterations, and a minimum lr of . 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 and RGB loss , as shown in Equation (6).
Here, supervises rendered depth with ground-truth depth, whereas enforces alignment with VFM-predicted depth. and are the and SSIM [51] losses between rendered and ground-truth images. We set , , , and 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.

Figure 5. Qualitative RGB novel-view synthesis results for test views.
| Method | Replica Dataset | TAT Dataset | RE10K Dataset | |||||||||||||||
| Input Views | Test Views | Input Views | Test Views | Input Views | Test Views | |||||||||||||
| P | S | L | P | S | L | P | S | L | P | S | L | P | S | L | P | S | L | |
| VolSplat | 16.15 | 0.667 | 0.471 | 15.44 | 0.636 | 0.467 | 15.45 | 0.595 | 0.419 | 16.33 | 0.733 | 0.241 | 20.55 | 0.731 | 0.415 | 14.13 | 0.552 | 0.546 |
| MVSplat | 19.97 | 0.759 | 0.318 | 19.41 | 0.736 | 0.319 | 16.48 | 0.615 | 0.380 | 15.05 | 0.522 | 0.460 | 14.72 | 0.648 | 0.441 | 14.64 | 0.611 | 0.488 |
| LongLRM | 27.94 | 0.882 | 0.212 | 23.39 | 0.839 | 0.276 | 22.01 | 0.705 | 0.302 | 12.18 | 0.415 | 0.600 | 26.12 | 0.893 | 0.206 | 17.44 | 0.814 | 0.322 |
| DA3 | 23.64 | 0.837 | 0.208 | 22.82 | 0.817 | 0.243 | 19.20 | 0.621 | 0.348 | 17.68 | 0.556 | 0.413 | 22.47 | 0.818 | 0.233 | 17.89 | 0.694 | 0.355 |
| VoxelTTO | 29.91 | 0.896 | 0.170 | 28.54 | 0.876 | 0.172 | 20.89 | 0.656 | 0.347 | 19.23 | 0.565 | 0.391 | 27.14 | 0.901 | 0.207 | 20.93 | 0.764 | 0.279 |
| AnySplat | 23.25 | 0.825 | 0.200 | 17.78 | 0.681 | 0.359 | 17.34 | 0.645 | 0.311 | 12.71 | 0.440 | 0.478 | 23.61 | 0.842 | 0.216 | 12.15 | 0.533 | 0.534 |
| YoNoSplat | 24.52 | 0.809 | 0.262 | 11.57 | 0.635 | 0.432 | 13.61 | 0.477 | 0.516 | 13.29 | 0.660 | 0.304 | 25.35 | 0.837 | 0.189 | 8.34 | 0.601 | 0.663 |
| MA | 14.35 | 0.579 | 0.533 | 17.95 | 0.757 | 0.227 | 12.66 | 0.497 | 0.570 | 14.66 | 0.592 | 0.381 | 13.76 | 0.583 | 0.589 | 10.01 | 0.430 | 0.673 |
| VoxelTTO (w/o TTO) | 29.28 | 0.890 | 0.174 | 25.83 | 0.827 | 0.196 | 23.74 | 0.767 | 0.247 | 14.71 | 0.485 | 0.475 | 26.97 | 0.893 | 0.216 | 17.46 | 0.673 | 0.378 |
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 . denote the set of pixels with valid ground-truth depth. For each pixel , and denote the rendered and ground-truth depths, respectively. Rel(%) is used to evaluate relative depth error: . Depth accuracy (%) is the proportion of pixels in satisfying . 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.

Figure 6. Qualitative depth novel-view synthesis results for test views.
| Method | Input Views | Test Views | ||||||||
| RMSE | Rel | RMSE | Rel | |||||||
| Replica Dataset | ||||||||||
| VolSplat | 102.8 | 30.42 | 11.88 | 22.55 | 46.78 | 202.2 | 63.89 | 0.812 | 2.443 | 6.859 |
| MVSplat | 216.0 | 69.26 | 1.984 | 3.894 | 9.261 | 1649 | 446.1 | 11.84 | 25.77 | 42.37 |
| LongLRM | 18.06 | 6.018 | 60.73 | 83.09 | 96.83 | 19.00 | 5.724 | 65.87 | 85.52 | 97.14 |
| DA3 | 125.5 | 43.27 | 28.30 | 46.40 | 57.85 | 172.0 | 38.69 | 64.18 | 71.19 | 76.29 |
| VoxelTTO | 6.167 | 0.688 | 98.66 | 99.29 | 99.72 | 6.945 | 0.866 | 98.86 | 99.24 | 99.58 |
| AnySplat | 33.72 | 9.582 | 59.05 | 76.88 | 86.14 | 35.92 | 10.78 | 45.43 | 67.09 | 89.86 |
| YoNoSplat | 23.47 | 7.281 | 51.07 | 75.58 | 94.78 | 242.1 | 78.73 | 0.304 | 0.692 | 2.643 |
| MA | 13.30 | 3.559 | 82.69 | 93.74 | 98.12 | 37.96 | 10.15 | 44.80 | 68.37 | 86.92 |
| VoxelTTO (w/o TTO) | 7.394 | 0.777 | 98.57 | 99.22 | 99.66 | 9.032 | 1.115 | 98.14 | 98.83 | 99.38 |
| TAT Dataset | ||||||||||
| VolSplat | 490.7 | 33.73 | 9.984 | 19.13 | 41.25 | 69.15 | 31.05 | 9.058 | 18.01 | 42.70 |
| MVSplat | 894.3 | 71.98 | 0.911 | 1.809 | 4.739 | 143.7 | 76.57 | 0.343 | 0.699 | 2.203 |
| LongLRM | 181.6 | 24.67 | 4.757 | 9.940 | 54.43 | 462.1 | 16.34 | 42.08 | 59.37 | 80.53 |
| DA3 | 456.9 | 41.44 | 18.76 | 36.71 | 60.34 | 299.4 | 30.49 | 3.934 | 8.181 | 23.29 |
| VoxelTTO | 90.63 | 2.875 | 90.12 | 95.83 | 98.15 | 46.22 | 3.235 | 86.75 | 95.16 | 98.33 |
| AnySplat | 117.4 | 4.793 | 70.01 | 90.98 | 96.95 | 499.1 | 360.3 | 5.405 | 15.37 | 30.40 |
| YoNoSplat | 511.5 | 37.06 | 8.018 | 15.69 | 33.54 | 83.57 | 47.59 | 7.480 | 15.38 | 38.70 |
| MA | 139.9 | 6.560 | 67.01 | 81.11 | 94.77 | 512.4 | 399.1 | 15.28 | 28.06 | 31.31 |
| VoxelTTO (w/o TTO) | 82.07 | 2.079 | 96.35 | 97.45 | 98.49 | 63.93 | 7.035 | 52.39 | 78.84 | 95.44 |
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: . 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.
| Method | Replica | TAT | ||
| Rot(°)↓ | Trans(m)↓ | Rot(°)↓ | Trans(m)↓ | |
| DA3 | 0.324 | 0.004 | 0.682 | 0.030 |
| VoxelTTO | 0.220 | 0.002 | 0.262 | 0.020 |
| AnySplat | 7.320 | 0.106 | 3.516 | 0.135 |
| YoNoSplat | 4.769 | 0.063 | 5.318 | 0.115 |
| MA | 4.438 | 0.056 | 3.232 | 0.065 |
| VoxelTTO (w/o TTO) | 0.301 | 0.006 | 1.636 | 0.047 |
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.
| Method | Acc.↓ | Com.↓ | Over.↓ |
| VoxelTTO (w/o TTO) | 5.556 | 9.159 | 7.357 |
| VoxelTTO (Camera Enc.) | 3.719 | 4.853 | 4.286 |
| VoxelTTO | 2.952 | 3.880 | 3.416 |
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 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.

Figure 7. Test-view RGB-D synthesis performance at different TTO iteration counts.
| Experiment | Setting | PSNR | SSIM | LPIPS | RMSE (cm) | Rel | |||
| Number of Gaussian Splits | 1 | 27.82 | 0.866 | 0.172 | 8.09 | 0.94 | 98.71 | 99.10 | 99.47 |
| Number of Gaussian Splits | 2 (Ours) | 28.54 | <u>0.876</u> | <u>0.172</u> | 6.95 | 0.87 | <u>98.86</u> | <u>99.24</u> | 99.58 |
| Number of Gaussian Splits | 4 | <u>28.35</u> | 0.876 | 0.179 | <u>7.21</u> | 0.86 | 98.93 | 99.24 | 99.54 |
| Stochastic Solid Volume Rendering | V-V (Ours) | 28.54 | 0.876 | 0.172 | 6.95 | 0.87 | 98.86 | 99.24 | 99.58 |
| Stochastic Solid Volume Rendering | V-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 Rendering | R-R | 21.52 | 0.774 | 0.269 | 27.58 | 4.05 | 87.76 | 95.11 | 97.10 |
| Stochastic Solid Volume Rendering | R-V | 22.37 | 0.798 | 0.260 | 15.82 | 3.04 | 89.32 | 97.03 | 98.66 |
| Voxel Size | 0.0015 | 29.34 | 0.891 | 0.142 | 6.62 | 0.83 | 99.04 | 99.36 | 99.64 |
| Voxel Size | 0.002 (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 Size | 0.003 | 26.78 | 0.856 | 0.211 | 9.45 | 1.18 | 98.32 | 98.88 | 99.32 |
| Voxel Size | 0.005 | 23.44 | 0.798 | 0.290 | 15.22 | 2.45 | 93.17 | 97.15 | 98.47 |
| Depth Loss | Ours | 28.54 | 0.876 | <u>0.172</u> | 6.95 | 0.87 | 98.86 | 99.24 | 99.58 |
| Depth Loss | No | 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 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

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 . 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 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 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 submanifold convolutions, combined with BatchNorm and ReLU. The 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 , scales , rotation quaternions , and opacity logits ; the appearance branch predicts 2-degree spherical harmonics (SH) coefficients. As each voxel splits into (default as 2) Gaussians, the output dimensions are , , , , and . For voxel anchor , the primitive’s center is calculated
is calculated as shown in Equation (7).
where is the center-offset head output, is a learned scalar initialized to , and . The center head is zero-initialized, so all Gaussians initially lie at the voxel anchor. Scales are mapped by a sigmoid to in the voxel coordinate system. Quaternions are normalized to unit length. Voxel confidence adjusts each opacity logit as shown in Equation (8).
The constant SH coefficients are initialized with the voxel RGB color by , 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 , where is the network output and 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 , the voxel size expressed in the ground-truth coordinate system becomes 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 between predicted and ground-truth input-camera centers using Umeyama alignment [41]. For world-to-camera extrinsics , the camera center is . We parameterize the transform as , where and . The optimal transform maps the prediction frame to the ground-truth frame, as shown in Equation (9).
Here, and denote predicted and ground-truth quantities. We define their cross-covariance matrix and the reflection-correction matrix as shown in Equation (10).
Here, is the singular value decomposition of . The matrix ensures that the recovered rotation has determinant . The parameters of have the closed-form solution, as shown in Equation (11).
The same transform is applied to the reconstructed scene and predicted cameras. For a Gaussian center , scale vector , and rotation matrix , the transform is shown in in Equation (12).
We also rotate the SH coefficients consistently with , as shown in Equation (13).
Here, 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 and compute , to be expressed in respective first-camera coordinate. is the translation of , the calculation of normalized translations is shown in Equation (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 , 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 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.
| Method | Metric | office0 | office1 | office2 | office3 | office4 | room0 | room1 | room2 | Average |
| VolSplat | PSNR↑ | 17.60 | 26.46 | 12.30 | 12.98 | 13.96 | 15.53 | 15.22 | 15.14 | 16.15 |
| VolSplat | SSIM↑ | 0.694 | 0.831 | 0.528 | 0.611 | 0.679 | 0.654 | 0.661 | 0.680 | 0.667 |
| VolSplat | LPIPS↓ | 0.436 | 0.285 | 0.578 | 0.491 | 0.483 | 0.516 | 0.479 | 0.499 | 0.471 |
| MVSplat | PSNR↑ | 21.46 | 24.83 | 17.66 | 17.85 | 17.67 | 20.55 | 19.32 | 20.40 | 19.97 |
| MVSplat | SSIM↑ | 0.792 | 0.782 | 0.667 | 0.730 | 0.774 | 0.812 | 0.747 | 0.768 | 0.759 |
| MVSplat | LPIPS↓ | 0.329 | 0.224 | 0.362 | 0.315 | 0.354 | 0.280 | 0.325 | 0.356 | 0.318 |
| AnySplat | PSNR↑ | 26.40 | 27.57 | 19.45 | 19.87 | 24.40 | 22.89 | 21.79 | 23.59 | 23.25 |
| AnySplat | SSIM↑ | 0.874 | 0.838 | 0.757 | 0.777 | 0.866 | 0.836 | 0.831 | 0.824 | 0.825 |
| AnySplat | LPIPS↓ | 0.165 | 0.194 | 0.227 | 0.243 | 0.192 | 0.193 | 0.174 | 0.215 | 0.200 |
| YoNoSplat | PSNR↑ | 25.12 | 29.12 | 22.09 | 24.36 | 25.91 | 22.83 | 22.24 | 24.49 | 24.52 |
| YoNoSplat | SSIM↑ | 0.807 | 0.882 | 0.753 | 0.823 | 0.852 | 0.805 | 0.750 | 0.803 | 0.809 |
| YoNoSplat | LPIPS↓ | 0.285 | 0.227 | 0.276 | 0.215 | 0.230 | 0.260 | 0.314 | 0.285 | 0.262 |
| DepthAnything3 | PSNR↑ | 24.51 | 28.66 | 21.80 | 22.40 | 24.27 | 23.73 | 20.74 | 23.01 | 23.64 |
| DepthAnything3 | SSIM↑ | 0.831 | 0.860 | 0.812 | 0.836 | 0.878 | 0.862 | 0.761 | 0.852 | 0.837 |
| DepthAnything3 | LPIPS↓ | 0.236 | 0.149 | 0.224 | 0.183 | 0.173 | 0.203 | 0.282 | 0.217 | 0.208 |
| MapAnything | PSNR↑ | 16.88 | 24.37 | 12.00 | 12.87 | 13.09 | 11.47 | 11.00 | 13.07 | 14.35 |
| MapAnything | SSIM↑ | 0.641 | 0.713 | 0.501 | 0.568 | 0.598 | 0.542 | 0.495 | 0.571 | 0.579 |
| MapAnything | LPIPS↓ | 0.494 | 0.349 | 0.572 | 0.552 | 0.549 | 0.566 | 0.626 | 0.557 | 0.533 |
| LongLRM | PSNR↑ | 26.20 | 29.39 | 27.32 | 27.63 | 29.35 | 28.45 | 26.02 | 29.15 | 27.94 |
| LongLRM | SSIM↑ | 0.852 | 0.845 | 0.863 | 0.891 | 0.923 | 0.914 | 0.864 | 0.904 | 0.882 |
| LongLRM | LPIPS↓ | 0.254 | 0.268 | 0.191 | 0.179 | 0.170 | 0.193 | 0.236 | 0.207 | 0.212 |
| VoxelTTO (w/o TTO) | PSNR↑ | 33.45 | 33.10 | 25.84 | 27.42 | 29.83 | 27.98 | 27.81 | 28.82 | 29.28 |
| VoxelTTO (w/o TTO) | SSIM↑ | 0.921 | 0.938 | 0.849 | 0.887 | 0.903 | 0.896 | 0.842 | 0.882 | 0.890 |
| VoxelTTO (w/o TTO) | LPIPS↓ | 0.138 | 0.155 | 0.188 | 0.143 | 0.180 | 0.182 | 0.204 | 0.199 | 0.174 |
| Ours | PSNR↑ | 33.44 | 32.78 | 27.20 | 26.97 | 30.67 | 28.88 | 28.90 | 30.47 | 29.91 |
| Ours | SSIM↑ | 0.920 | 0.930 | 0.866 | 0.879 | 0.913 | 0.907 | 0.850 | 0.901 | 0.896 |
| Ours | LPIPS↓ | 0.142 | 0.162 | 0.183 | 0.158 | 0.170 | 0.171 | 0.192 | 0.186 | 0.170 |
Table A1. Per-scene RGB view synthesis results on Replica input views.
| Method | Metric | office0 | office1 | office2 | office3 | office4 | room0 | room1 | room2 | Average |
| VolSplat | PSNR↑ | 16.70 | 26.13 | 11.67 | 11.50 | 13.06 | 14.88 | 14.69 | 14.87 | 15.44 |
| VolSplat | SSIM↑ | 0.640 | 0.857 | 0.488 | 0.538 | 0.632 | 0.624 | 0.657 | 0.653 | 0.636 |
| VolSplat | LPIPS↓ | 0.450 | 0.249 | 0.579 | 0.501 | 0.513 | 0.493 | 0.453 | 0.496 | 0.467 |
| MVSplat | PSNR↑ | 20.51 | 23.61 | 16.74 | 17.61 | 16.46 | 20.99 | 19.88 | 19.50 | 19.41 |
| MVSplat | SSIM↑ | 0.766 | 0.740 | 0.640 | 0.695 | 0.747 | 0.802 | 0.767 | 0.731 | 0.736 |
| MVSplat | LPIPS↓ | 0.272 | 0.254 | 0.370 | 0.336 | 0.387 | 0.293 | 0.249 | 0.392 | 0.319 |
| AnySplat | PSNR↑ | 15.99 | 22.74 | 15.31 | 15.38 | 18.46 | 19.75 | 16.46 | 18.19 | 17.78 |
| AnySplat | SSIM↑ | 0.666 | 0.732 | 0.595 | 0.603 | 0.750 | 0.731 | 0.686 | 0.686 | 0.681 |
| AnySplat | LPIPS↓ | 0.387 | 0.325 | 0.417 | 0.383 | 0.326 | 0.310 | 0.355 | 0.369 | 0.359 |
| YoNoSplat | PSNR↑ | 12.11 | 17.43 | 8.71 | 9.33 | 10.64 | 11.17 | 10.94 | 12.24 | 11.57 |
| YoNoSplat | SSIM↑ | 0.597 | 0.530 | 0.594 | 0.618 | 0.691 | 0.700 | 0.661 | 0.689 | 0.635 |
| YoNoSplat | LPIPS↓ | 0.407 | 0.386 | 0.486 | 0.455 | 0.375 | 0.417 | 0.474 | 0.453 | 0.432 |
| DepthAnything3 | PSNR↑ | 24.41 | 27.67 | 20.29 | 21.72 | 22.92 | 23.76 | 19.25 | 22.57 | 22.82 |
| DepthAnything3 | SSIM↑ | 0.800 | 0.837 | 0.788 | 0.824 | 0.856 | 0.863 | 0.714 | 0.849 | 0.817 |
| DepthAnything3 | LPIPS↓ | 0.250 | 0.205 | 0.256 | 0.224 | 0.220 | 0.238 | 0.312 | 0.242 | 0.243 |
| MapAnything | PSNR↑ | 20.69 | 24.37 | 15.37 | 14.81 | 15.53 | 17.29 | 17.09 | 18.49 | 17.95 |
| MapAnything | SSIM↑ | 0.810 | 0.772 | 0.688 | 0.682 | 0.753 | 0.778 | 0.799 | 0.776 | 0.757 |
| MapAnything | LPIPS↓ | 0.195 | 0.267 | 0.254 | 0.254 | 0.262 | 0.208 | 0.167 | 0.212 | 0.227 |
| LongLRM | PSNR↑ | 27.03 | 29.03 | 21.96 | 21.46 | 22.78 | 22.48 | 17.83 | 24.55 | 23.39 |
| LongLRM | SSIM↑ | 0.844 | 0.892 | 0.816 | 0.826 | 0.875 | 0.864 | 0.741 | 0.857 | 0.839 |
| LongLRM | LPIPS↓ | 0.289 | 0.268 | 0.281 | 0.249 | 0.248 | 0.257 | 0.348 | 0.265 | 0.276 |
| VoxelTTO (w/o TTO) | PSNR↑ | 26.30 | 29.10 | 20.89 | 25.41 | 29.09 | 25.84 | 23.59 | 26.46 | 25.83 |
| VoxelTTO (w/o TTO) | SSIM↑ | 0.799 | 0.883 | 0.737 | 0.851 | 0.899 | 0.863 | 0.742 | 0.838 | 0.827 |
| VoxelTTO (w/o TTO) | LPIPS↓ | 0.209 | 0.179 | 0.232 | 0.148 | 0.181 | 0.191 | 0.223 | 0.204 | 0.196 |
| Ours | PSNR↑ | 29.56 | 27.99 | 24.87 | 27.90 | 30.75 | 28.88 | 28.24 | 30.14 | 28.54 |
| Ours | SSIM↑ | 0.857 | 0.876 | 0.818 | 0.892 | 0.920 | 0.905 | 0.833 | 0.910 | 0.876 |
| Ours | LPIPS↓ | 0.168 | 0.188 | 0.198 | 0.148 | 0.166 | 0.167 | 0.171 | 0.169 | 0.172 |
Table A2. Per-scene RGB view synthesis results on Replica test views.
| Method | Metric | office0 | office1 | office2 | office3 | office4 | room0 | room1 | room2 | Average |
| VolSplat | MAE (cm) | 99.52 | 22.56 | 91.10 | 104.1 | 98.42 | 84.86 | 96.17 | 71.38 | 83.52 |
| VolSplat | RMSE (cm) | 122.2 | 27.92 | 115.9 | 128.9 | 117.5 | 109.2 | 115.8 | 85.66 | 102.8 |
| VolSplat | Rel (%) | 41.23 | 12.38 | 28.14 | 30.70 | 34.18 | 24.25 | 41.00 | 31.53 | 30.42 |
| VolSplat | (%) | 8.301 | 25.93 | 11.24 | 9.267 | 8.674 | 14.17 | 8.788 | 8.687 | 11.88 |
| VolSplat | (%) | 15.58 | 48.36 | 21.89 | 18.06 | 16.57 | 25.84 | 16.73 | 17.44 | 22.55 |
| VolSplat | (%) | 32.39 | 85.10 | 46.70 | 41.78 | 36.50 | 55.39 | 35.10 | 41.36 | 46.78 |
| MVSplat | MAE (cm) | 133.1 | 146.1 | 281.7 | 223.5 | 246.9 | 138.8 | 155.6 | 195.8 | 190.2 |
| MVSplat | RMSE (cm) | 151.3 | 161.3 | 303.8 | 263.4 | 275.6 | 184.6 | 172.9 | 215.7 | 216.0 |
| MVSplat | Rel (%) | 56.75 | 72.56 | 86.47 | 63.69 | 81.58 | 42.90 | 67.61 | 82.59 | 69.26 |
| MVSplat | (%) | 4.369 | 2.738 | 0.6278 | 1.962 | 0.7422 | 2.527 | 2.038 | 0.8712 | 1.984 |
| MVSplat | (%) | 8.597 | 5.229 | 1.225 | 3.847 | 1.506 | 5.028 | 4.031 | 1.689 | 3.894 |
| MVSplat | (%) | 18.97 | 11.46 | 2.886 | 9.234 | 3.852 | 13.47 | 10.21 | 4.018 | 9.261 |
| AnySplat | MAE (cm) | 15.33 | 12.49 | 16.72 | 10.66 | 13.08 | 7.317 | 102.7 | 10.35 | 23.58 |
| AnySplat | RMSE (cm) | 21.24 | 15.14 | 24.47 | 18.00 | 23.90 | 12.73 | 133.5 | 20.92 | 33.72 |
| AnySplat | Rel (%) | 6.711 | 6.857 | 5.349 | 3.169 | 4.697 | 2.293 | 43.26 | 4.319 | 9.582 |
| AnySplat | (%) | 48.66 | 47.27 | 53.36 | 84.03 | 67.75 | 92.50 | 0.05539 | 78.81 | 59.05 |
| AnySplat | (%) | 73.81 | 74.47 | 91.58 | 95.34 | 88.47 | 97.23 | 0.1207 | 94.02 | 76.88 |
| AnySplat | (%) | 98.62 | 99.10 | 98.00 | 98.72 | 98.37 | 98.70 | 0.5146 | 97.11 | 86.14 |
| YoNoSplat | MAE (cm) | 31.29 | 16.74 | 13.27 | 14.78 | 18.96 | 12.28 | 19.57 | 13.63 | 17.56 |
| YoNoSplat | RMSE (cm) | 37.73 | 20.77 | 21.03 | 19.42 | 26.07 | 17.60 | 25.04 | 20.13 | 23.47 |
| YoNoSplat | Rel (%) | 14.27 | 8.795 | 4.301 | 4.679 | 6.814 | 4.019 | 8.881 | 6.494 | 7.281 |
| YoNoSplat | (%) | 20.88 | 35.42 | 75.51 | 63.57 | 50.44 | 73.46 | 35.08 | 54.24 | 51.07 |
| YoNoSplat | (%) | 40.76 | 61.90 | 90.21 | 91.96 | 83.12 | 91.35 | 63.78 | 81.59 | 75.58 |
| YoNoSplat | (%) | 79.73 | 94.17 | 98.08 | 99.31 | 96.59 | 99.11 | 94.74 | 96.53 | 94.78 |
| DepthAnything3 | MAE (cm) | 232.9 | 26.78 | 23.22 | 346.4 | 19.04 | 15.76 | 232.1 | 20.04 | 114.5 |
| DepthAnything3 | RMSE (cm) | 236.9 | 34.98 | 42.54 | 361.3 | 29.90 | 26.60 | 240.1 | 31.98 | 125.5 |
| DepthAnything3 | Rel (%) | 100.00 | 14.22 | 8.333 | 100.00 | 7.436 | 5.337 | 100.0 | 10.91 | 43.27 |
| DepthAnything3 | (%) | 0 | 21.52 | 44.52 | 0.004174 | 53.25 | 72.70 | 0 | 34.47 | 28.30 |
| DepthAnything3 | (%) | 0 | 35.04 | 80.94 | 0.004174 | 84.81 | 90.68 | 0 | 79.75 | 46.40 |
| DepthAnything3 | (%) | 0 | 85.16 | 93.90 | 0.004174 | 94.85 | 96.98 | 0 | 91.96 | 57.85 |
| MapAnything | MAE (cm) | 8.206 | 4.766 | 8.696 | 9.216 | 8.335 | 6.638 | 8.647 | 8.772 | 7.909 |
| MapAnything | RMSE (cm) | 11.08 | 6.849 | 15.59 | 14.96 | 17.48 | 12.25 | 12.45 | 15.76 | 13.30 |
| MapAnything | Rel (%) | 3.604 | 2.612 | 3.086 | 3.373 | 3.785 | 2.657 | 4.106 | 5.254 | 3.559 |
| MapAnything | (%) | 75.89 | 84.24 | 86.13 | 84.86 | 87.68 | 91.86 | 68.64 | 82.25 | 82.69 |
| MapAnything | (%) | 95.31 | 97.36 | 93.08 | 94.93 | 92.43 | 96.04 | 92.98 | 87.86 | 93.74 |
| MapAnything | (%) | 99.72 | 99.74 | 98.63 | 98.36 | 96.35 | 98.78 | 99.11 | 94.29 | 98.12 |
| LongLRM | MAE (cm) | 14.73 | 11.70 | 12.99 | 18.67 | 11.58 | 10.63 | 17.62 | 11.27 | 13.65 |
| LongLRM | RMSE (cm) | 18.30 | 14.64 | 17.99 | 23.03 | 17.24 | 13.94 | 23.75 | 15.61 | 18.06 |
| LongLRM | Rel (%) | 6.568 | 6.190 | 4.485 | 6.352 | 4.618 | 3.683 | 9.001 | 7.253 | 6.018 |
| LongLRM | (%) | 43.68 | 50.14 | 71.00 | 52.53 | 76.59 | 77.18 | 45.82 | 68.95 | 60.73 |
| LongLRM | (%) | 76.28 | 79.62 | 89.98 | 82.60 | 89.40 | 93.67 | 71.01 | 82.24 | 83.09 |
| LongLRM | (%) | 99.18 | 98.53 | 98.81 | 98.12 | 97.58 | 99.31 | 91.36 | 91.77 | 96.83 |
| VoxelTTO (w/o TTO) | MAE (cm) | 1.353 | 1.530 | 2.728 | 2.728 | 2.551 | 2.075 | 1.779 | 1.969 | 2.089 |
| VoxelTTO (w/o TTO) | RMSE (cm) | 4.066 | 8.787 | 9.599 | 6.773 | 10.64 | 5.222 | 4.782 | 9.292 | 7.394 |
| VoxelTTO (w/o TTO) | Rel (%) | 0.5837 | 0.8316 | 0.8506 | 0.8495 | 0.8632 | 0.6281 | 0.7987 | 0.8169 | 0.777 |
| VoxelTTO (w/o TTO) | (%) | 98.77 | 98.79 | 97.92 | 98.34 | 98.71 | 99.03 | 99.18 | 97.87 | 98.57 |
| VoxelTTO (w/o TTO) | (%) | 99.44 | 99.42 | 98.79 | 99.17 | 99.08 | 99.50 | 99.61 | 98.78 | 99.22 |
| VoxelTTO (w/o TTO) | (%) | 99.85 | 99.92 | 99.52 | 99.69 | 99.41 | 99.81 | 99.83 | 99.29 | 99.66 |
| Ours | MAE (cm) | 1.420 | 1.040 | 2.564 | 2.538 | 2.093 | 1.873 | 1.706 | 1.493 | 1.841 |
| Ours | RMSE (cm) | 4.392 | 3.014 | 9.278 | 6.468 | 8.644 | 5.419 | 4.640 | 7.484 | 6.167 |
| Ours | Rel (%) | 0.6103 | 0.5425 | 0.8111 | 0.8096 | 0.7150 | 0.5802 | 0.8050 | 0.6374 | 0.688 |
| Ours | (%) | 98.76 | 98.78 | 97.99 | 98.18 | 99.06 | 99.05 | 99.22 | 98.31 | 98.66 |
| Ours | (%) | 99.41 | 99.37 | 98.86 | 99.20 | 99.35 | 99.47 | 99.62 | 99.07 | 99.29 |
| Ours | (%) | 99.88 | 99.92 | 99.58 | 99.70 | 99.60 | 99.79 | 99.83 | 99.48 | 99.72 |
Table A3. Per-scene depth view synthesis results on Replica input views.
| Method | Metric | office0 | office1 | office2 | office3 | office4 | room0 | room1 | room2 | Average |
| VolSplat | MAE (cm) | 168.2 | 45.77 | 225.5 | 259.5 | 209.1 | 247.9 | 187.0 | 175.8 | 189.9 |
| VolSplat | RMSE (cm) | 175.5 | 52.01 | 242.9 | 275.1 | 226.3 | 260.0 | 196.3 | 190.3 | 202.2 |
| VolSplat | Rel (%) | 70.96 | 22.52 | 69.42 | 69.98 | 65.03 | 69.38 | 75.71 | 68.15 | 63.89 |
| VolSplat | (%) | 0.2459 | 5.606 | 0.001700 | 0.01207 | 0.1346 | 0.001524 | 0.04609 | 0.4491 | 0.812 |
| VolSplat | (%) | 0.6678 | 17.01 | 0.01927 | 0.05175 | 0.6836 | 0.01554 | 0.2028 | 0.8918 | 2.443 |
| VolSplat | (%) | 2.155 | 44.84 | 0.2042 | 0.1729 | 4.029 | 0.2157 | 1.101 | 2.153 | 6.859 |
| MVSplat | MAE (cm) | 2982 | 1531 | 278.3 | 166.1 | 288.6 | 71.78 | 1951 | 249.9 | 939.7 |
| MVSplat | RMSE (cm) | 4284 | 3178 | 787.3 | 370.5 | 681.0 | 175.9 | 3139 | 584.8 | 1649 |
| MVSplat | Rel (%) | 1372 | 847.4 | 103.2 | 75.61 | 135.8 | 24.46 | 872.5 | 138.3 | 446.1 |
| MVSplat | (%) | 0.4064 | 16.65 | 14.21 | 14.50 | 5.539 | 27.82 | 4.695 | 10.90 | 11.84 |
| MVSplat | (%) | 0.9160 | 35.41 | 32.40 | 35.43 | 16.38 | 51.72 | 8.942 | 25.03 | 25.77 |
| MVSplat | (%) | 2.851 | 48.72 | 59.11 | 60.99 | 30.45 | 75.09 | 19.13 | 42.67 | 42.37 |
| AnySplat | MAE (cm) | 42.23 | 17.02 | 19.62 | 22.94 | 17.57 | 14.20 | 45.11 | 21.91 | 25.08 |
| AnySplat | RMSE (cm) | 51.27 | 22.43 | 33.71 | 34.62 | 30.70 | 20.52 | 55.33 | 38.80 | 35.92 |
| AnySplat | Rel (%) | 18.55 | 8.242 | 7.137 | 8.378 | 6.889 | 4.803 | 20.96 | 11.32 | 10.78 |
| AnySplat | (%) | 15.90 | 32.46 | 54.70 | 62.92 | 60.78 | 79.71 | 10.93 | 46.10 | 45.43 |
| AnySplat | (%) | 31.21 | 64.18 | 79.30 | 76.76 | 90.39 | 94.83 | 25.37 | 74.74 | 67.09 |
| AnySplat | (%) | 71.46 | 94.96 | 93.39 | 94.02 | 97.65 | 97.45 | 79.14 | 90.85 | 89.86 |
| YoNoSplat | MAE (cm) | 173.0 | 153.7 | 262.2 | 266.0 | 248.3 | 293.8 | 212.4 | 205.8 | 226.9 |
| YoNoSplat | RMSE (cm) | 180.5 | 160.1 | 282.2 | 292.6 | 267.6 | 309.2 | 222.7 | 222.5 | 242.1 |
| YoNoSplat | Rel (%) | 72.73 | 76.51 | 78.30 | 74.41 | 76.09 | 83.25 | 87.94 | 80.62 | 78.73 |
| YoNoSplat | (%) | 1.807 | 0.02397 | 0.05995 | 0.2666 | 0.04517 | 0.01748 | 0 | 0.2114 | 0.304 |
| YoNoSplat | (%) | 3.486 | 0.05690 | 0.2673 | 1.019 | 0.2036 | 0.06027 | 0.001422 | 0.4477 | 0.692 |
| YoNoSplat | (%) | 6.891 | 0.3992 | 3.360 | 6.029 | 1.567 | 0.8187 | 0.2642 | 1.816 | 2.643 |
| DepthAnything3 | MAE (cm) | 291.6 | 39.52 | 10.19 | 16.87 | 12.66 | 6.408 | 375.9 | 10.95 | 95.51 |
| DepthAnything3 | RMSE (cm) | 539.3 | 70.33 | 42.61 | 142.4 | 35.93 | 16.49 | 495.3 | 34.27 | 172.0 |
| DepthAnything3 | Rel (%) | 116.1 | 18.69 | 4.328 | 6.751 | 7.132 | 2.477 | 146.5 | 7.549 | 38.69 |
| DepthAnything3 | (%) | 17.04 | 44.40 | 91.56 | 83.90 | 92.18 | 95.66 | 0.2916 | 88.41 | 64.18 |
| DepthAnything3 | (%) | 38.14 | 57.96 | 94.32 | 90.70 | 95.13 | 97.04 | 2.990 | 93.26 | 71.19 |
| DepthAnything3 | (%) | 47.66 | 67.97 | 95.94 | 96.82 | 96.14 | 97.96 | 12.66 | 95.21 | 76.29 |
| MapAnything | MAE (cm) | 37.45 | 15.31 | 14.78 | 34.39 | 14.66 | 19.17 | 54.47 | 18.53 | 26.09 |
| MapAnything | RMSE (cm) | 46.05 | 19.82 | 30.37 | 45.80 | 28.09 | 26.65 | 69.74 | 37.16 | 37.96 |
| MapAnything | Rel (%) | 15.10 | 7.580 | 5.145 | 10.32 | 5.860 | 6.193 | 21.49 | 9.509 | 10.15 |
| MapAnything | (%) | 8.761 | 37.05 | 81.34 | 32.74 | 69.79 | 52.51 | 8.289 | 67.96 | 44.80 |
| MapAnything | (%) | 34.27 | 71.44 | 90.38 | 64.58 | 85.95 | 88.73 | 23.74 | 87.89 | 68.37 |
| MapAnything | (%) | 73.23 | 96.12 | 94.80 | 92.36 | 97.28 | 96.09 | 54.17 | 91.37 | 86.92 |
| LongLRM | MAE (cm) | 14.25 | 11.01 | 10.04 | 17.70 | 11.55 | 9.825 | 19.16 | 11.15 | 13.09 |
| LongLRM | RMSE (cm) | 18.20 | 14.30 | 16.23 | 25.53 | 20.20 | 14.53 | 25.58 | 17.48 | 19.00 |
| LongLRM | Rel (%) | 6.281 | 5.910 | 3.428 | 6.528 | 4.530 | 3.272 | 8.829 | 7.022 | 5.724 |
| LongLRM | (%) | 46.18 | 55.56 | 82.17 | 62.60 | 80.27 | 85.09 | 44.24 | 70.89 | 65.87 |
| LongLRM | (%) | 80.16 | 80.68 | 95.13 | 87.48 | 91.85 | 96.24 | 67.98 | 84.65 | 85.52 |
| LongLRM | (%) | 98.85 | 98.42 | 99.12 | 97.59 | 97.89 | 99.04 | 92.81 | 93.46 | 97.14 |
| VoxelTTO (w/o TTO) | MAE (cm) | 2.947 | 1.440 | 4.297 | 3.128 | 1.774 | 1.615 | 2.782 | 2.932 | 2.615 |
| VoxelTTO (w/o TTO) | RMSE (cm) | 7.395 | 3.849 | 15.06 | 10.08 | 11.39 | 5.572 | 7.083 | 11.83 | 9.032 |
| VoxelTTO (w/o TTO) | Rel (%) | 1.299 | 0.7562 | 1.411 | 1.475 | 0.6557 | 0.5455 | 1.173 | 1.611 | 1.115 |
| VoxelTTO (w/o TTO) | (%) | 97.00 | 98.54 | 97.49 | 98.28 | 99.38 | 99.29 | 97.66 | 97.52 | 98.14 |
| VoxelTTO (w/o TTO) | (%) | 98.10 | 99.10 | 98.29 | 98.61 | 99.44 | 99.50 | 99.30 | 98.31 | 98.83 |
| VoxelTTO (w/o TTO) | (%) | 99.38 | 99.83 | 98.84 | 99.12 | 99.56 | 99.69 | 99.74 | 98.93 | 99.38 |
| Ours | MAE (cm) | 1.940 | 1.279 | 3.422 | 2.281 | 2.686 | 1.192 | 1.959 | 1.356 | 2.014 |
| Ours | RMSE (cm) | 5.766 | 4.082 | 11.69 | 8.005 | 8.360 | 4.847 | 4.275 | 8.536 | 6.945 |
| Ours | Rel (%) | 0.8473 | 0.6823 | 1.143 | 1.156 | 0.9632 | 0.3954 | 0.8587 | 0.8849 | 0.866 |
| Ours | (%) | 98.24 | 98.44 | 98.43 | 98.77 | 99.61 | 99.47 | 99.29 | 98.71 | 98.86 |
| Ours | (%) | 98.90 | 99.01 | 98.92 | 99.01 | 99.68 | 99.64 | 99.73 | 99.07 | 99.24 |
| Ours | (%) | 99.63 | 99.76 | 99.28 | 99.30 | 99.76 | 99.74 | 99.91 | 99.31 | 99.58 |
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 by minimizing the squared error over , which contains the pixels with valid depth values in the ground-truth depth map.

Figure A1. Additional qualitative results for novel view synthesis.
Here, and are the rendered and ground-truth depths at pixel . We compute depth metrics using on valid pixels . For evaluation on Tanks and Temples (TAT), we use the

Figure A2. Additional qualitative results for depth view synthesis.
| Scan | VoxelTTO (w/o TTO) | VoxelTTO (Camera Enc.) | VoxelTTO | ||||||
| Acc.↓ | Com.↓ | Over.↓ | Acc.↓ | Com.↓ | Over.↓ | Acc.↓ | Com.↓ | Over.↓ | |
| scan24 | 4.741 | 4.601 | 4.671 | 1.552 | 2.434 | 1.993 | 3.483 | 4.945 | 4.214 |
| scan37 | 11.175 | 10.745 | 10.960 | 9.224 | 10.200 | 9.712 | 5.086 | 6.636 | 5.861 |
| scan40 | 3.227 | 5.456 | 4.341 | 1.998 | 3.226 | 2.612 | 0.721 | 2.141 | 1.431 |
| scan55 | 6.045 | 9.953 | 7.999 | 2.063 | 2.291 | 2.177 | 1.395 | 2.425 | 1.910 |
| scan63 | 5.356 | 7.236 | 6.296 | 4.687 | 6.079 | 5.383 | 2.393 | 2.418 | 2.405 |
| scan65 | 4.457 | 9.713 | 7.085 | 2.451 | 3.383 | 2.917 | 2.044 | 2.267 | 2.155 |
| scan69 | 4.526 | 4.273 | 4.400 | 5.358 | 5.604 | 5.481 | 2.273 | 2.866 | 2.569 |
| scan83 | 9.504 | 13.363 | 11.433 | 5.330 | 11.255 | 8.293 | 4.677 | 6.899 | 5.788 |
| scan97 | 5.561 | 11.483 | 8.522 | 5.108 | 8.424 | 6.766 | 4.958 | 6.678 | 5.818 |
| scan105 | 7.929 | 15.648 | 11.789 | 3.753 | 4.648 | 4.200 | 1.374 | 2.052 | 1.713 |
| scan106 | 4.357 | 10.619 | 7.488 | 2.968 | 4.125 | 3.547 | 3.702 | 4.981 | 4.342 |
| scan110 | 3.292 | 3.865 | 3.579 | 2.202 | 1.897 | 2.050 | 2.418 | 3.035 | 2.726 |
| scan114 | 4.861 | 14.501 | 9.681 | 2.196 | 2.325 | 2.260 | 3.951 | 4.291 | 4.121 |
| scan118 | 4.022 | 8.643 | 6.332 | 3.480 | 3.490 | 3.485 | 2.790 | 3.361 | 3.075 |
| scan122 | 4.282 | 7.283 | 5.782 | 3.416 | 3.407 | 3.411 | 3.010 | 3.205 | 3.107 |
| Average | 5.556 | 9.159 | 7.357 | 3.719 | 4.853 | 4.286 | 2.952 | 3.880 | 3.416 |
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.
| Method | Metric | Barn | Church | Courthouse | Average |
| VolSplat | 14.15 | 16.09 | 16.12 | 15.45 | |
| 0.543 | 0.538 | 0.704 | 0.595 | ||
| 0.431 | 0.474 | 0.351 | 0.419 | ||
| MVSplat | 15.37 | 16.34 | 17.73 | 16.48 | |
| 0.594 | 0.534 | 0.716 | 0.615 | ||
| 0.410 | 0.424 | 0.306 | 0.380 | ||
| AnySplat | 15.81 | 18.53 | 17.67 | 17.34 | |
| 0.646 | 0.580 | 0.708 | 0.645 | ||
| 0.296 | 0.318 | 0.320 | 0.311 | ||
| YoNoSplat | 12.62 | 14.10 | 14.11 | 13.61 | |
| 0.475 | 0.346 | 0.609 | 0.477 | ||
| 0.501 | 0.590 | 0.457 | 0.516 | ||
| DepthAnything3 | 19.61 | 17.46 | 20.53 | 19.20 | |
| 0.628 | 0.503 | 0.732 | 0.621 | ||
| 0.332 | 0.405 | 0.307 | 0.348 | ||
| MapAnything | 9.98 | 14.41 | 13.59 | 12.66 | |
| 0.460 | 0.411 | 0.621 | 0.497 | ||
| 0.593 | 0.632 | 0.486 | 0.570 | ||
| LongLRM | 21.82 | 20.66 | 23.55 | 22.01 | |
| 0.695 | 0.611 | 0.809 | 0.705 | ||
| 0.279 | 0.372 | 0.254 | 0.302 | ||
| VoxelTTO (w/o TTO) | 22.30 | 22.03 | 26.90 | 23.74 | |
| 0.725 | 0.707 | 0.869 | 0.767 | ||
| 0.278 | 0.274 | 0.189 | 0.247 | ||
| Ours | 20.38 | 18.01 | 24.30 | 20.89 | |
| 0.634 | 0.520 | 0.813 | 0.656 | ||
| 0.355 | 0.432 | 0.254 | 0.347 |
Table A5. Per-scene RGB view synthesis results on TAT input views.
Table A6. Per-scene RGB view synthesis results on TAT test views.
| Method | Metric | Barn | Church | Courthouse | Average |
| VolSplat | 15.92 | 15.97 | 17.10 | 16.33 | |
| 0.786 | 0.573 | 0.840 | 0.733 | ||
| 0.181 | 0.365 | 0.177 | 0.241 | ||
| MVSplat | 13.60 | 16.09 | 15.47 | 15.05 | |
| 0.495 | 0.423 | 0.649 | 0.522 | ||
| 0.477 | 0.504 | 0.400 | 0.460 | ||
| AnySplat | 11.67 | 14.63 | 11.81 | 12.71 | |
| 0.405 | 0.365 | 0.550 | 0.440 | ||
| 0.440 | 0.526 | 0.469 | 0.478 | ||
| YoNoSplat | 12.17 | 12.30 | 15.41 | 13.29 | |
| 0.668 | 0.491 | 0.822 | 0.660 | ||
| 0.300 | 0.456 | 0.155 | 0.304 | ||
| DepthAnything3 | 17.22 | 16.59 | 19.23 | 17.68 | |
| 0.553 | 0.411 | 0.704 | 0.556 | ||
| 0.404 | 0.509 | 0.325 | 0.413 | ||
| MapAnything | 13.68 | 15.34 | 14.97 | 14.66 | |
| 0.551 | 0.568 | 0.658 | 0.592 | ||
| 0.447 | 0.370 | 0.327 | 0.381 | ||
| LongLRM | 10.46 | 15.03 | 11.05 | 12.18 | |
| 0.383 | 0.330 | 0.530 | 0.415 | ||
| 0.588 | 0.614 | 0.599 | 0.600 | ||
| VoxelTTO (w/o TTO) | 13.67 | 14.93 | 15.53 | 14.71 | |
| 0.453 | 0.341 | 0.661 | 0.485 | ||
| 0.477 | 0.560 | 0.388 | 0.475 | ||
| Ours | 19.31 | 16.16 | 22.21 | 19.23 | |
| 0.561 | 0.369 | 0.765 | 0.565 | ||
| 0.378 | 0.510 | 0.286 | 0.391 |
Table A6. Per-scene RGB view synthesis results on TAT test views.
| Method | Metric | Barn | Church | Courthouse | Average |
| VolSplat | MAE (cm) | 289.1 | 350.8 | 472.0 | 370.7 |
| VolSplat | RMSE (cm) | 378.2 | 423.5 | 670.4 | 490.7 |
| VolSplat | Rel (%) | 31.84 | 39.59 | 29.79 | 33.73 |
| VolSplat | (%) | 11.06 | 7.643 | 11.25 | 9.984 |
| VolSplat | (%) | 21.33 | 14.75 | 21.34 | 19.13 |
| VolSplat | (%) | 45.67 | 31.89 | 46.20 | 41.25 |
| MVSplat | MAE (cm) | 536.6 | 619.2 | 1214 | 789.9 |
| MVSplat | RMSE (cm) | 661.7 | 663.4 | 1358 | 894.3 |
| MVSplat | Rel (%) | 61.40 | 72.12 | 82.42 | 71.98 |
| MVSplat | (%) | 0.9033 | 1.575 | 0.2568 | 0.911 |
| MVSplat | (%) | 1.791 | 3.131 | 0.5058 | 1.809 |
| MVSplat | (%) | 5.010 | 7.991 | 1.218 | 4.739 |
| AnySplat | MAE (cm) | 42.72 | 32.20 | 91.59 | 55.50 |
| AnySplat | RMSE (cm) | 66.90 | 50.98 | 234.5 | 117.4 |
| AnySplat | Rel (%) | 4.822 | 4.107 | 5.450 | 4.793 |
| AnySplat | (%) | 61.75 | 73.73 | 74.57 | 70.01 |
| AnySplat | (%) | 93.38 | 90.38 | 89.19 | 90.98 |
| AnySplat | (%) | 97.98 | 98.69 | 94.21 | 96.95 |
| YoNoSplat | MAE (cm) | 306.1 | 135.6 | 856.8 | 432.8 |
| YoNoSplat | RMSE (cm) | 369.3 | 183.6 | 981.8 | 511.5 |
| YoNoSplat | Rel (%) | 34.52 | 16.33 | 60.34 | 37.06 |
| YoNoSplat | (%) | 3.963 | 19.12 | 0.9748 | 8.018 |
| YoNoSplat | (%) | 7.114 | 37.94 | 2.037 | 15.69 |
| YoNoSplat | (%) | 17.26 | 78.13 | 5.246 | 33.54 |
| DepthAnything3 | MAE (cm) | 110.0 | 830.5 | 173.5 | 371.3 |
| DepthAnything3 | RMSE (cm) | 183.9 | 853.5 | 333.4 | 456.9 |
| DepthAnything3 | Rel (%) | 12.61 | 100.00 | 11.74 | 41.44 |
| DepthAnything3 | (%) | 22.18 | 0 | 34.11 | 18.76 |
| DepthAnything3 | (%) | 43.97 | 0 | 66.18 | 36.71 |
| DepthAnything3 | (%) | 92.41 | 0 | 88.62 | 60.34 |
| MapAnything | MAE (cm) | 102.2 | 20.27 | 89.87 | 70.77 |
| MapAnything | RMSE (cm) | 161.4 | 30.48 | 227.8 | 139.9 |
| MapAnything | Rel (%) | 11.36 | 2.502 | 5.821 | 6.560 |
| MapAnything | (%) | 29.16 | 90.90 | 80.98 | 67.01 |
| MapAnything | (%) | 60.01 | 97.15 | 86.19 | 81.11 |
| MapAnything | (%) | 92.02 | 99.58 | 92.73 | 94.77 |
| LongLRM | MAE (cm) | 150.4 | 160.8 | 198.3 | 169.8 |
| LongLRM | RMSE (cm) | 157.2 | 187.5 | 200.2 | 181.6 |
| LongLRM | Rel (%) | 21.80 | 25.82 | 26.41 | 24.67 |
| LongLRM | (%) | 0.7080 | 12.07 | 1.496 | 4.757 |
| LongLRM | (%) | 3.746 | 23.83 | 2.244 | 9.940 |
| LongLRM | (%) | 69.75 | 53.19 | 40.36 | 54.43 |
| VoxelTTO (w/o TTO) | MAE (cm) | 17.80 | 8.689 | 48.00 | 24.83 |
| VoxelTTO (w/o TTO) | RMSE (cm) | 66.76 | 21.58 | 157.9 | 82.07 |
| VoxelTTO (w/o TTO) | Rel (%) | 2.051 | 1.174 | 3.012 | 2.079 |
| VoxelTTO (w/o TTO) | (%) | 97.49 | 96.76 | 94.81 | 96.35 |
| VoxelTTO (w/o TTO) | (%) | 98.32 | 98.55 | 95.48 | 97.45 |
| VoxelTTO (w/o TTO) | (%) | 99.42 | 99.58 | 96.47 | 98.49 |
| Ours | MAE (cm) | 21.72 | 20.13 | 53.68 | 31.84 |
| Ours | RMSE (cm) | 77.56 | 36.63 | 157.7 | 90.63 |
| Ours | Rel (%) | 2.318 | 2.660 | 3.649 | 2.875 |
| Ours | (%) | 94.43 | 86.50 | 89.45 | 90.12 |
| Ours | (%) | 97.66 | 94.89 | 94.95 | 95.83 |
| Ours | (%) | 99.24 | 98.80 | 96.43 | 98.15 |
Table A7. Per-scene depth view synthesis results on TAT input views.
| Method | Metric | Barn | Church | Courthouse | Average |
| VolSplat | MAE (cm) | 34.48 | 55.18 | 66.27 | 51.98 |
| VolSplat | RMSE (cm) | 41.87 | 68.75 | 96.84 | 69.15 |
| VolSplat | Rel (%) | 25.58 | 41.08 | 26.51 | 31.05 |
| VolSplat | (%) | 10.34 | 6.365 | 10.47 | 9.058 |
| VolSplat | (%) | 20.38 | 11.97 | 21.69 | 18.01 |
| VolSplat | (%) | 47.32 | 26.15 | 54.63 | 42.70 |
| MVSplat | MAE (cm) | 100.3 | 105.6 | 171.3 | 125.7 |
| MVSplat | RMSE (cm) | 117.2 | 111.3 | 202.7 | 143.7 |
| MVSplat | Rel (%) | 74.26 | 81.12 | 74.35 | 76.57 |
| MVSplat | (%) | 0.05092 | 0.5351 | 0.4454 | 0.343 |
| MVSplat | (%) | 0.1420 | 1.077 | 0.8783 | 0.699 |
| MVSplat | (%) | 0.6206 | 3.890 | 2.099 | 2.203 |
| AnySplat | MAE (cm) | 638.6 | 737.5 | 31.71 | 469.3 |
| AnySplat | RMSE (cm) | 685.4 | 754.7 | 57.45 | 499.1 |
| AnySplat | Rel (%) | 484.4 | 584.2 | 12.42 | 360.3 |
| AnySplat | (%) | 0 | 0 | 16.22 | 5.405 |
| AnySplat | (%) | 0 | 0 | 46.12 | 15.37 |
| AnySplat | (%) | 0 | 0 | 91.23 | 30.40 |
| YoNoSplat | MAE (cm) | 52.33 | 94.52 | 66.88 | 71.24 |
| YoNoSplat | RMSE (cm) | 62.11 | 100.5 | 88.09 | 83.57 |
| YoNoSplat | Rel (%) | 40.03 | 73.62 | 29.13 | 47.59 |
| YoNoSplat | (%) | 8.938 | 2.013 | 11.49 | 7.480 |
| YoNoSplat | (%) | 19.14 | 4.177 | 22.83 | 15.38 |
| YoNoSplat | (%) | 48.24 | 10.38 | 57.51 | 38.70 |
| DepthAnything3 | MAE (cm) | 351.9 | 160.1 | 70.36 | 194.1 |
| DepthAnything3 | RMSE (cm) | 512.4 | 271.5 | 114.5 | 299.4 |
| DepthAnything3 | Rel (%) | 39.93 | 20.09 | 31.48 | 30.49 |
| DepthAnything3 | (%) | 0.8630 | 10.68 | 0.2642 | 3.934 |
| DepthAnything3 | (%) | 1.581 | 22.38 | 0.5835 | 8.181 |
| DepthAnything3 | (%) | 3.096 | 64.93 | 1.852 | 23.29 |
| MapAnything | MAE (cm) | 716.8 | 744.5 | 21.43 | 494.2 |
| MapAnything | RMSE (cm) | 729.9 | 757.7 | 49.87 | 512.4 |
| MapAnything | Rel (%) | 548.1 | 641.1 | 8.352 | 399.1 |
| MapAnything | (%) | 0 | 0 | 45.86 | 15.28 |
| MapAnything | (%) | 0 | 0 | 84.19 | 28.06 |
| MapAnything | (%) | 0 | 0 | 93.95 | 31.31 |
| LongLRM | MAE (cm) | 81.74 | 176.9 | 266.0 | 174.9 |
| LongLRM | RMSE (cm) | 284.7 | 224.5 | 877.0 | 462.1 |
| LongLRM | Rel (%) | 9.165 | 23.07 | 16.80 | 16.34 |
| LongLRM | (%) | 57.86 | 18.27 | 50.14 | 42.08 |
| LongLRM | (%) | 78.90 | 32.74 | 66.47 | 59.37 |
| LongLRM | (%) | 93.52 | 59.91 | 88.17 | 80.53 |
| VoxelTTO (w/o TTO) | MAE (cm) | 58.24 | 46.64 | 19.20 | 41.36 |
| VoxelTTO (w/o TTO) | RMSE (cm) | 86.94 | 65.80 | 39.06 | 63.93 |
| VoxelTTO (w/o TTO) | Rel (%) | 6.607 | 6.744 | 7.756 | 7.035 |
| VoxelTTO (w/o TTO) | (%) | 52.79 | 67.39 | 37.01 | 52.39 |
| VoxelTTO (w/o TTO) | (%) | 76.68 | 80.81 | 79.03 | 78.84 |
| VoxelTTO (w/o TTO) | (%) | 94.22 | 96.40 | 95.71 | 95.44 |
| Ours | MAE (cm) | 24.96 | 25.92 | 7.919 | 19.60 |
| Ours | RMSE (cm) | 69.75 | 45.91 | 23.01 | 46.22 |
| Ours | Rel (%) | 2.782 | 3.438 | 3.488 | 3.235 |
| Ours | (%) | 90.99 | 84.04 | 85.24 | 86.75 |
| Ours | (%) | 97.11 | 91.74 | 96.63 | 95.16 |
| Ours | (%) | 99.40 | 98.14 | 97.44 | 98.33 |
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.
| Method | Metric | 89999a83cb78ec83 | 2a7918811fcee38d | 0cc7cbbf | 749f321b | 42fb0ecb | b90b641a12a0a776 | 0e3fc05a | Average | |||
| VolSplat | PSNR↑ | 16.68 | 22.61 | 21.75 | 21.88 | 16.61 | 19.16 | 19.77 | 18.80 | 24.05 | 24.16 | 20.55 |
| SSIM↑ | 0.594 | 0.819 | 0.804 | 0.817 | 0.583 | 0.639 | 0.741 | 0.662 | 0.806 | 0.841 | 0.731 | |
| LPIPS↓ | 0.458 | 0.394 | 0.369 | 0.381 | 0.533 | 0.521 | 0.423 | 0.426 | 0.293 | 0.350 | 0.415 | |
| MVSplat | PSNR↑ | 11.24 | 14.53 | 16.73 | 12.14 | 14.93 | 11.41 | 18.46 | 12.52 | 15.52 | 19.77 | 14.72 |
| SSIM↑ | 0.494 | 0.723 | 0.717 | 0.727 | 0.571 | 0.551 | 0.744 | 0.488 | 0.626 | 0.839 | 0.648 | |
| LPIPS↓ | 0.548 | 0.427 | 0.376 | 0.416 | 0.499 | 0.530 | 0.418 | 0.491 | 0.409 | 0.296 | 0.441 | |
| AnySplat | PSNR↑ | 19.50 | 21.03 | 19.64 | 32.65 | 21.85 | 25.14 | 19.94 | 24.23 | 27.17 | 24.95 | 23.61 |
| SSIM↑ | 0.760 | 0.852 | 0.766 | 0.962 | 0.797 | 0.872 | 0.768 | 0.859 | 0.878 | 0.904 | 0.842 | |
| LPIPS↓ | 0.223 | 0.246 | 0.307 | 0.131 | 0.231 | 0.188 | 0.328 | 0.138 | 0.152 | 0.219 | 0.216 | |
| YoNoSplat | PSNR↑ | 22.06 | 30.25 | 25.75 | 32.03 | 22.51 | 20.00 | 26.84 | 22.46 | 22.29 | 29.28 | 25.35 |
| SSIM↑ | 0.808 | 0.947 | 0.885 | 0.944 | 0.776 | 0.695 | 0.887 | 0.781 | 0.735 | 0.909 | 0.837 | |
| LPIPS↓ | 0.184 | 0.146 | 0.173 | 0.139 | 0.217 | 0.311 | 0.166 | 0.165 | 0.218 | 0.168 | 0.189 | |
| DepthAnything3 | PSNR↑ | 19.11 | 24.65 | 20.36 | 25.94 | 17.56 | 24.06 | 24.34 | 22.71 | 21.89 | 24.02 | 22.47 |
| SSIM↑ | 0.780 | 0.920 | 0.801 | 0.918 | 0.646 | 0.854 | 0.859 | 0.815 | 0.715 | 0.865 | 0.818 | |
| LPIPS↓ | 0.217 | 0.185 | 0.277 | 0.183 | 0.368 | 0.199 | 0.193 | 0.165 | 0.281 | 0.263 | 0.233 | |
| MapAnything | PSNR↑ | 12.26 | 11.96 | 15.86 | 11.63 | 15.29 | 13.37 | 15.73 | 14.58 | 14.46 | 12.46 | 13.76 |
| SSIM↑ | 0.492 | 0.572 | 0.655 | 0.538 | 0.602 | 0.534 | 0.665 | 0.570 | 0.633 | 0.571 | 0.583 | |
| LPIPS↓ | 0.581 | 0.623 | 0.562 | 0.627 | 0.530 | 0.645 | 0.563 | 0.544 | 0.542 | 0.670 | 0.589 | |
| LongLRM | PSNR↑ | 24.59 | 29.53 | 25.53 | 29.75 | 23.15 | 26.54 | 23.80 | 24.37 | 25.40 | 28.54 | 26.12 |
| SSIM↑ | 0.883 | 0.949 | 0.908 | 0.944 | 0.847 | 0.907 | 0.839 | 0.835 | 0.882 | 0.933 | 0.893 | |
| LPIPS↓ | 0.171 | 0.182 | 0.212 | 0.187 | 0.242 | 0.189 | 0.259 | 0.183 | 0.208 | 0.225 | 0.206 | |
| VoxelTTO (w/o TTO) | PSNR↑ | 22.16 | 27.99 | 23.72 | 31.05 | 25.92 | 26.28 | 26.45 | 25.69 | 28.91 | 31.51 | 26.97 |
| SSIM↑ | 0.832 | 0.929 | 0.851 | 0.952 | 0.871 | 0.881 | 0.886 | 0.877 | 0.902 | 0.949 | 0.893 | |
| LPIPS↓ | 0.229 | 0.232 | 0.280 | 0.164 | 0.225 | 0.239 | 0.236 | 0.194 | 0.180 | 0.177 | 0.216 | |
| Ours | PSNR↑ | 22.28 | 28.12 | 25.74 | 30.34 | 26.12 | 26.28 | 26.25 | 26.10 | 29.38 | 30.76 | 27.14 |
| SSIM↑ | 0.844 | 0.939 | 0.905 | 0.943 | 0.876 | 0.882 | 0.884 | 0.886 | 0.913 | 0.942 | 0.901 | |
| LPIPS↓ | 0.225 | 0.222 | 0.224 | 0.171 | 0.214 | 0.244 | 0.236 | 0.189 | 0.163 | 0.183 | 0.207 |
Table A9. Per-scene RGB view synthesis results on RE10K input views.
| Method | Metric | 89999a83cb78ec83 | 2a7918811fcee38d | 0cc7cbbf | 749f321b | 42fb0ecb | b90b641a12a0a776 | 0e3fc05a | Average | |||
| VolSplat | PSNR↑ | 11.41 | 14.88 | 13.64 | 14.43 | 12.35 | 12.77 | 17.21 | 14.10 | 14.55 | 15.93 | 14.13 |
| SSIM↑ | 0.423 | 0.637 | 0.588 | 0.608 | 0.407 | 0.459 | 0.624 | 0.506 | 0.583 | 0.683 | 0.552 | |
| LPIPS↓ | 0.570 | 0.535 | 0.520 | 0.532 | 0.636 | 0.620 | 0.523 | 0.552 | 0.493 | 0.481 | 0.546 | |
| MVSplat | PSNR↑ | 11.33 | 14.72 | 16.69 | 12.58 | 14.73 | 11.35 | 16.50 | 12.76 | 15.41 | 20.31 | 14.64 |
| SSIM↑ | 0.466 | 0.692 | 0.687 | 0.719 | 0.523 | 0.494 | 0.667 | 0.442 | 0.589 | 0.830 | 0.611 | |
| LPIPS↓ | 0.577 | 0.472 | 0.420 | 0.421 | 0.543 | 0.579 | 0.548 | 0.528 | 0.450 | 0.340 | 0.488 | |
| AnySplat | PSNR↑ | 10.58 | 9.54 | 13.63 | 12.35 | 11.51 | 9.47 | 14.48 | 12.43 | 15.28 | 12.20 | 12.15 |
| SSIM↑ | 0.385 | 0.575 | 0.567 | 0.676 | 0.435 | 0.429 | 0.571 | 0.469 | 0.562 | 0.663 | 0.533 | |
| LPIPS↓ | 0.607 | 0.516 | 0.529 | 0.471 | 0.622 | 0.591 | 0.546 | 0.509 | 0.420 | 0.528 | 0.534 | |
| YoNoSplat | PSNR↑ | 8.42 | 8.90 | 7.85 | 10.21 | 8.32 | 7.98 | 9.14 | 6.07 | 6.90 | 9.60 | 8.34 |
| SSIM↑ | 0.508 | 0.714 | 0.547 | 0.781 | 0.506 | 0.600 | 0.611 | 0.464 | 0.499 | 0.774 | 0.601 | |
| LPIPS↓ | 0.702 | 0.590 | 0.647 | 0.602 | 0.695 | 0.673 | 0.683 | 0.714 | 0.701 | 0.624 | 0.663 | |
| DepthAnything3 | PSNR↑ | 15.98 | 20.89 | 19.79 | 14.41 | 17.03 | 17.66 | 20.04 | 16.90 | 19.68 | 16.54 | 17.89 |
| SSIM↑ | 0.583 | 0.826 | 0.773 | 0.728 | 0.619 | 0.635 | 0.738 | 0.604 | 0.650 | 0.785 | 0.694 | |
| LPIPS↓ | 0.317 | 0.248 | 0.300 | 0.511 | 0.392 | 0.392 | 0.302 | 0.302 | 0.342 | 0.445 | 0.355 | |
| MapAnything | PSNR↑ | 11.37 | 9.98 | 12.16 | 9.05 | 10.40 | 9.19 | 12.39 | 10.21 | 11.10 | 4.23 | 10.01 |
| SSIM↑ | 0.409 | 0.526 | 0.511 | 0.515 | 0.378 | 0.374 | 0.535 | 0.362 | 0.403 | 0.290 | 0.430 | |
| LPIPS↓ | 0.625 | 0.687 | 0.640 | 0.691 | 0.697 | 0.706 | 0.646 | 0.674 | 0.661 | 0.706 | 0.673 | |
| LongLRM | PSNR↑ | 15.24 | 14.90 | 16.86 | 19.72 | 18.91 | 18.47 | 20.13 | 20.35 | 15.29 | 14.59 | 17.44 |
| SSIM↑ | 0.767 | 0.841 | 0.825 | 0.888 | 0.794 | 0.813 | 0.790 | 0.798 | 0.810 | 0.815 | 0.814 | |
| LPIPS↓ | 0.313 | 0.377 | 0.306 | 0.308 | 0.302 | 0.325 | 0.321 | 0.241 | 0.265 | 0.460 | 0.322 | |
| VoxelTTO (w/o TTO) | PSNR↑ | 13.51 | 19.24 | 16.61 | 15.28 | 19.45 | 16.60 | 19.50 | 17.75 | 17.41 | 19.28 | 17.46 |
| SSIM↑ | 0.495 | 0.778 | 0.644 | 0.737 | 0.688 | 0.617 | 0.712 | 0.630 | 0.634 | 0.799 | 0.673 | |
| LPIPS↓ | 0.440 | 0.324 | 0.445 | 0.461 | 0.308 | 0.442 | 0.352 | 0.305 | 0.357 | 0.347 | 0.378 | |
| Ours | PSNR↑ | 17.64 | 22.89 | 19.54 | 19.90 | 21.18 | 20.13 | 22.68 | 21.11 | 22.46 | 21.81 | 20.93 |
| SSIM↑ | 0.679 | 0.860 | 0.730 | 0.801 | 0.735 | 0.695 | 0.793 | 0.762 | 0.765 | 0.824 | 0.764 | |
| LPIPS↓ | 0.281 | 0.266 | 0.312 | 0.345 | 0.260 | 0.322 | 0.289 | 0.225 | 0.213 | 0.277 | 0.279 |
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]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]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]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]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]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]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]Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022. 3
- [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]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]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]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]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]Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 2
- [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]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]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]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]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]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]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]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]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]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]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]Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2arxiv.org/abs/2406.09756
- [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]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]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]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]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]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]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]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]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]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]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]Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2DOI
- [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]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]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]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]Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2arxiv.org/abs/2408.16061
- [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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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]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