acceptodds
Withdrawn from ICLR 2027

VoxelTTO: Voxel-Aligned Feed-Forward 3D Gaussian Splatting with Test-Time Optimization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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 scenes from an arbitrary number of images and optional camera parameters. VoxelTTO aggregates dense image features into a global voxel representation 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, and DTU demonstrate improved RGB-D NVS and camera-pose estimation relative to prior methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Visual Foundation Models, Novel View Synthesis, Test-Time Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_danAOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes VoxelTTO, a feed-forward framework for reconstructing geometrically accurate 3DGS scenes from an arbitrary number of images and optional camera parameters. Instead of predicting pixel-aligned Gaussians, it aggregates dense image features into a global voxel representation and decodes multiple Gaussian primitives from each voxel, aiming to reduce overlap and improve geometric consistency. The paper further introduces stochastic solid volume rendering for geometry-aware supervision and a test-time optimization (TTO) strategy that inserts lightweight LoRA modules into a frozen VFM backbone to refine camera predictions when camera priors are available. Experiments on Replica, Tanks and Temples, and DTU demonstrate improvements in RGB/depth novel-view synthesis and camera pose estimation over previous feed-forward methods.

Critical strengths

This paper addresses a key limitation of pixel-aligned feed-forward 3DGS. The voxel-aligned representation is well motivated and provides a principled way to reduce overlapping Gaussian primitives while allowing multiple Gaussians to model complex local geometry.

Critical weaknesses
  • The manuscript would benefit from careful proofreading. Many citations are improperly typeset, with author names and years directly attached to the text instead of being enclosed in parentheses (e.g., through standard \citep commands), which noticeably reduces readability. The paper also contains numerous formatting issues, including missing spaces, missing punctuation, and overly crowded tables.
  • The experimental evaluation is not entirely convincing. Rather than evaluating on widely adopted benchmarks such as RE10K, DL3DV, and ACID, which are used by many recent feed-forward baselines, the paper instead focuses on Replica and Tanks and Temples. Including results on the commonly used benchmarks would better position the method with respect to the existing literature.
Questions for authors

Mentioned above.

Reviewer_Rx7HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents VoxelTTO, a voxel-aligned feed-forward 3D Gaussian Splatting framework that reconstructs scenes from arbitrary image collections with optional camera parameters. It aggregates features from a frozen visual foundation model into a global voxel representation and decodes Gaussian primitives using a sparse convolutional U-Net and a self-splitting MLP. It further introduces pose-supervised test-time adaptation of lightweight LoRA modules to exploit available camera priors, and uses stochastic solid volume rendering to improve geometric fidelity. Experiments on Replica, Tanks and Temples, and DTU report improvements over prior feed-forward methods in camera-pose estimation and RGB-D rendering for both input and novel views.

Critical strengths

The paper effectively combines voxel-aligned Gaussian reconstruction with pose-supervised test-time optimization (TTO). The experimental results demonstrate that integrating these components substantially improves camera-pose estimation and RGB-D rendering, highlighting the practical value of the approach.

Critical weaknesses
  1. Limited novelty: The proposed architecture closely resembles that of VolSplat. Both construct feature voxels from predicted depth maps and processed multi-view features, refine the voxel features using a 3D U-Net, and then decode Gaussian primitives through a Gaussian prediction head. The main difference is the introduction of a test-time loss, which is similar in spirit to the evaluation-time pose alignment (EPA) strategy in NoPoSplat, where poses are optimized at test time by minimizing photometric error to improve alignment with ground-truth poses. The proposed method therefore appears to be primarily a combination of existing techniques, with limited methodological novelty.
  2. Insufficiently controlled experimental comparisons: All reported results are obtained under cross-dataset evaluation, yet the paper does not establish whether the competing methods were trained on the same data. Consequently, differences in training data may confound the comparisons, and the reported performance gains cannot be conclusively attributed to the proposed method.
Questions for authors
  1. The approach appears conceptually close to existing methods, particularly VolSplat’s voxel-aligned reconstruction and NoPoSplat’s test-time pose alignment, without an obvious new insight beyond their integration.
  2. What training datasets and pretrained checkpoints were used for each baseline, and how do they compare with those used by VoxelTTO? Are there controlled comparisons under matched training conditions that support attributing the reported gains to the proposed method rather than differences in training data or pretraining?
  3. Most figures and tables have only brief titles, with little explanatory detail in their captions, making them difficult to interpret.
Reviewer_HJV9OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a voxel-based feed-forward model to reconstruct Gaussian splats from multiple images, tying the Gaussian count to the number of occupied voxels rather than image pixels. The model uses DA3 to predict a point cloud with features, which is voxelized into sparse voxel latents. A sparse 3D U-Net refines these latents before predicting per-voxel Gaussians. A TTO module incorporates known camera information by optimizing LoRA parameters to align the predicted cameras with the known cameras. This is shown to be more effective than DA3’s original camera-conditioning strategy. The experimental results are promising.

Critical strengths

Predicting Gaussians from voxel features is well motivated, and the overall pipeline is well designed. The TTO strategy for incorporating known camera information into a feed-forward model is interesting and shows clear improvements over standard camera conditioning.

Critical weaknesses

The main comparison tables (Tab 1–3) are underspecified. Methods using known camera parameters should be grouped separately from those operating on unposed images. Access to known cameras provides a substantial advantage, so results from these two settings should not be directly compared without qualification. For the known-camera setting, more advanced methods such as DepthSplat (CVPR'25) and LongLRM (ICCV'25) are missing from the comparison.

Questions for authors
  1. How well does the predefined voxel size generalize across scenes of different scales? How does it affect feed-forward inference and rasterization runtime? Can the voxel size be adjusted at inference time to trade off efficiency and quality?
  2. Fig. 2: It would be clearer to use different arrow styles for training-time and test-time backpropagation, distinguishing updates to the reconstruction modules during training from LoRA updates during TTO.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.