acceptodds
Under review as a conference paper at ICLR 2027

InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views

12 views

Abstract

Feed-forward Gaussian reconstruction builds 3D scenes directly from input images for novel view rendering, but learning Gaussian attributes from predicted geometry for high-quality rendering under large viewpoint changes remains challenging. We analyze limitations in Gaussian attribute learning and introduce InfiniSplat v2, which improves rendering quality through architectural and training changes. Our method decouples geometry and attribute prediction with a frozen geometry branch, aligns predicted Gaussians using ground-truth camera poses during training, and adopts large-angle target view supervision. Across six datasets, the feed-forward outputs of the same trained model achieve state-of-the-art mean LPIPS and DISTS among the evaluated methods in both pose-conditioned and pose-free settings. In the pose-conditioned setting, mean LPIPS and DISTS decrease by 25.8% and 32.9%, respectively, compared with the strongest evaluated baselines. We further find that opacity learned with only image reconstruction supervision suppresses geometrically inconsistent points.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3DGS

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TKzZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

InfiniSplat v2 studies Gaussian attribute learning on top of a pretrained geometry model. It uses a frozen DA3 branch for depth and camera prediction and a separate, partially fine-tuned branch for Gaussian attributes. Gaussian centers remain fixed, while training uses scale-aligned ground-truth cameras and target sampling biased toward larger viewing-direction differences. An optional learned refiner updates attributes using input-view rendering gradients. The same model is evaluated on six datasets in pose-conditioned and pose-free settings, with substantial improvements in mean LPIPS and DISTS. The paper also analyzes the relationship between learned opacity and geometric inconsistency.

Critical strengths
  1. Substantial perceptual improvements across multiple settings. Table 1 reports clear improvements over the strongest corresponding baselines across six datasets. The feed-forward output reduces mean LPIPS and DISTS by 25.8% and 32.9% in the pose-conditioned setting, with substantial gains also observed without pose inputs. The same trained model supports both settings, and the main improvements do not depend on refinement.

  2. A useful design supported by controlled ablations. Table 2 and Appendix B.2 isolate attribute-backbone adaptation, geometry freezing, positional offsets, camera handling, and target-view sampling while retaining the other training settings. In particular, the frozen-attribute-backbone control provides a meaningful matched-training comparison. These experiments support the practical value of preserving the geometric prior while adapting Gaussian attributes.

Critical weaknesses
  1. The evaluation protocol is not sufficiently specified to verify the comparisons. The paper does not clearly report the evaluated scene and example counts, exact splits and view-selection procedure, input/rendering/scoring resolutions, or baseline checkpoints and configurations. The source and coverage of the evaluation valid regions also need clarification. The training resolutions in Appendix A.1 and latency resolutions in Table 9 do not establish the settings used for the main comparisons. These omissions limit reproducibility and confidence in the reported margins. The practical cost is also incompletely characterized: Table 9 measures neural-module latency rather than complete reconstruction and refinement time.

  2. The analysis supports an effective recipe more strongly than it supports the proposed mechanisms. The geometry and offset ablations demonstrate worse rendering metrics under the chosen training setup, but quantitative geometric errors are needed to substantiate the interpretation that position updates compensate for attribute errors. The large-angle explanation is also under-tested. Equation (5) measures optical-axis differences, which need not correspond to increased parallax or different observation rays for the same surface. Figure 8 shows a mean angle increase from 12.34° to 17.97°, but no evaluation stratified by actual baseline or viewing difficulty is provided. This limits the strength of the paper’s analytical contribution and its claims about extreme viewpoints.

Questions for authors
  1. Evaluation clarification: Please specify the evaluated scenes and example counts, context/target sampling, input/rendering/scoring resolutions, baseline checkpoints and configurations, and evaluation-mask construction. Are the different input-view counts evaluated on fixed targets with nested context sets?

  2. Evidence for the interpretation: Using the existing evaluation scenes and ablated models, could you report quantitative geometry errors for the full model, w/o G, and w/ O, together with novel-view performance grouped by translation baseline or common-surface viewing-ray angle? This would help distinguish the proposed geometric explanation from other training effects and establish where the sampling strategy provides its gains.

  3. Computational cost: Please report complete reconstruction latency, peak memory, Gaussian counts, and the additional cost of the three-step refiner, with comparisons to the main baselines under compatible hardware and resolution settings.

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

This paper proposes a feed-forward Gaussian Splatting method, called InfiniteSplat V2. The proposed method utilizes the Depth Anything 3 as the geometry part, fully using of the previous methods and providing a avenue for others to follow. They also provide a simple and effective training paradigm, which is easy to understand and follow. The experiments show the proposed is supiror than some previous baselines.

Critical strengths
  1. This paper proposes a feed-forward GS method for fast and high-quality 3d reconstruction.

  2. The proposed method is simple and easy to follow.

  3. The language is well-organized and easy to follow.

  4. The experiments are conducted well. The authors report results on six datasets.

Critical weaknesses
  1. The title is vacuous. I did not get any highlights from the title. In my opinion, the title should highlight the most important part of the paper.

  2. I am curious why the authors did not compare with the v1 version of InfinitSplat. The authors should explain this since the method is called v2. I think it is necessary to compare with v1 version at least.

  3. The experiments only conduct with the number of views 4, 8, 16. I would like to know whether the proposed method still dominates in metrics if the number of views continues to grow.

  4. To Eq.6, why there's a stop gradient sign. Why in this design? There's no reasonale explanation. In my opinion, the training should be end-to-end so that the model can learn better. It seems the authors seprately train the proposed method into different modules. I don't get the idea why doing in this way.

  5. Actually, the proposed method is quite simple, and it seems there are no nay new insights from this paper. The design is straightforward, although the performance is good.

  6. I am curious about why the authors only present LPISP and DISTS metrics. In general, PSNR and SSIM are also very common for novel view synthesis. They may provide some indication of which method is better.

  7. There's no benchmark for self-driving tasks and baselines (quantitatively). It would be very helpful to highlight this method if the authors can conduct it.

Questions for authors
  1. Will the authors release the source code and pretrained weights? I believe it would be very helpful for the industry and acamedia.

  2. I suggest the authors create a project page for better demonstration.

Reviewer_DhF4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a feed-forward 3DGS method that improves results over previous methods on six datasets. However, it reads more like a good implementation than a research paper. The main challenge is not clearly defined, and it is unclear how the proposed method addresses it. Each component is a standard design choice, and the paper mainly shows good numbers.

Critical strengths

The improvement is large. In the pose-conditioned setting, mean LPIPS and DISTS drop by 25.8% and 32.9% relative to the strongest baseline, and the visual results are clearly sharper (Table 1, Figures 3 and 4).

Critical weaknesses
  1. The novelty is not clear. A frozen geometry branch, GT-pose alignment, and large-angle target-view supervision are all basic design choices in feed-forward 3DGS. Related work such as QuerySplat (frozen geometric features) and DA3-GS (Gaussian attributes on top of a geometry model) already uses similar ideas. The paper does not explain what is new beyond combining them, or why they overcome the stated challenge. The ablations cover only two datasets, so it is hard to tell how much each component really contributes.

  2. Figure 1 shows a noisy point cloud but a clean 3DGS rendering. Neither the caption nor the text around the figure explains why. The authors emphasize the good results but never tell the reader how the noisy points are handled in this example. Opacity suppression is discussed elsewhere, but it is never linked to Figure 1, so the reader has to guess. It is also unclear whether this is an intended design or a side observation.

Questions for authors

The method is named InfiniSplat v2, but v1 is never clearly introduced. What is v1, what exactly changes in v2, and how much of the improvement comes from the change? v1 is also not compared in the experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.