acceptodds
Under review as a conference paper at ICLR 2027

Dual-Splatting for Efficient UHD 3D Gaussian Splatting Super-Resolution

Abstract

3D Gaussian Splatting (3DGS) enables novel view synthesis with high visual quality, but its rendering quality remains limited by the resolution of captured observations. To support close inspection, recent 3DGS super-resolution methods introduce high-frequency priors from pretrained single image super-resolution (SISR) models. However, fitting generated RGB details can require dense Gaussian primitives and high resolution rasterization, limiting ultra high definition (UHD) scalability on consumer hardware. To address this issue, we propose Dual-Splatting, a framework that integrates the SISR latent generative process into a dual branch representation. The RGB branch reconstructs scene structure and appearance from low resolution observations, while the latent branch models high-frequency details through SISR latent priors. The two branches share geometry to align scene reconstruction and detail synthesis. A latent refinement head transforms features rendered at the input resolution into latents for the frozen SISR decoder, enabling detail synthesis without Gaussian rasterization at the target resolution. To further improve consistency across views, we introduce multiview regularization in latent space that aggregates reprojected predictions from neighboring views. Experiments across multiple datasets and resolutions show compact scene representations and improved consistency over independent SISR, with practical 4K/6K super-resolution on a 24 GB GPU.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Super Resolution, Latent Space, Ultra-High-Definition

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

This paper proposes Dual-Splatting, a 3DGS super-resolution framework with an RGB branch and a latent branch. The RGB branch reconstructs scene structure from LR observations, while the latent branch models high-frequency details using frozen SISR priors. The two branches share Gaussian geometry, and a latent refinement head decodes rendered features into HR images. Experiments show reduced Gaussian count and memory usage, supporting 4K/6K synthesis on a 24 GB GPU.

Critical strengths

The proposed method performs Gaussian splatting super-resolution in latent space. The motivation is reasonable and the method is sound. The innovation lies in leveraging SISR at the latent level to achieve super-resolution, which is a sensible way to reduce Gaussian count and memory usage for UHD synthesis.

Critical weaknesses

The main concern is that both the latent generation process and the decoder of the SISR model are frozen. This means the upper bound of the method is determined by the SISR prior. The SISR model used here is InvSR, which is diffusion-based and therefore has some hallucination. The evaluation metrics for the super-resolved renderings are all perceptual quality metrics, which cannot rule out the influence of hallucination. As a result, the fidelity of the synthesized HR details remains unverified.

Questions for authors

The paper states that PSNR/SSIM are measured at the captured scale, but it is not entirely clear what this scale is in each setting. Could the authors clarify the exact resolution at which LR fidelity is measured for the 1K and 2K target settings? If the dataset provides 1K ground truth, could a fidelity comparison of the SR output at 1K resolution also be reported?

Since the perceptual quality of the super-resolution depends heavily on the SISR model, and both Figure 3 and the MUSIQ scores in Table 3 indicate that the results are visually very close, why does MANIQA show a relatively large gap? Could the authors explain this discrepancy?

Reviewer_5BhGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses UHD 3D Gaussian Splatting super-resolution on consumer GPUs. Its motivation is that fitting SISR-generated details directly into Gaussian RGB appearance can require more primitives and rasterization at the target resolution, making high-resolution training expensive.

Dual-Splatting uses an RGB branch to reconstruct structure and appearance from low-resolution observations and a latent branch to represent details from a frozen SISR model, InvSR. The branches share Gaussian geometry: positions are anchor centers plus offsets, while covariance and opacity are jointly predicted from RGB features and latent codes. Both branches are rasterized at low resolution with shared compositing weights. A refinement head with high-frequency, low-frequency, and residual stems maps the rendered features to a four-channel latent representation, which is normalized and decoded into a high-resolution image. Multiview regularization reprojects neighboring latent predictions using rendered depth and aggregates them with depth, direction, and latent-appearance consistency weights. Experiments cover T&T/DB, Mip-NeRF 360, and Mill 19/UrbanScene3D from 1K to 6K, reporting captured-scale fidelity, no-reference perceptual quality, cross-view reprojection error (CVRE), Gaussian count, and peak training memory.

Critical strengths

1. UHD synthesis under a limited memory budget has clear practical value. Rasterizing at low resolution and decoding to high resolution allows the method to reach 4K/6K on a 24 GB GPU, where the four evaluated 3DGSSR baselines encounter training OOM. This provides a feasible route to UHD super-resolution. The reported Gaussian counts are also substantially lower; for example, the method uses 0.85M primitives at 2K on Mip-NeRF 360, compared with 4.26M for SRGS.

2. The approach targets a relevant limitation of independent SISR: inconsistency across views. Sharing geometry between RGB and latent representations is intended to constrain variations in synthesized detail. In Table 3, relative to independent 3DGS-InvSR at 4K, CVRE L1 decreases from 0.0546 to 0.0427 and reprojection PSNR increases from 21.01 to 23.25 dB. This suggests a potentially useful consistency–perceptual-quality tradeoff, although the combined benefit still needs confirmation under a fixed configuration.

Critical weaknesses

1. Selecting each metric independently across configurations prevents assessment of an actual quality–resource operating point and weakens the ablations. Section 4.2 explicitly states that each metric is selected independently across configurations, and Figure 1 applies this selection to both quality and memory. Table 4 further acknowledges that the full-model and no-RGB-branch results are not obtained under matched settings. Thus, reported quality, memory, and Gaussian count may not describe a single runnable model, making it impossible to determine whether the claimed quality and efficiency advantages hold simultaneously. This directly affects the central contribution of UHD super-resolution under resource constraints.

Likewise, the CVRE degradation without the RGB branch cannot be attributed entirely to that branch under unmatched settings. The different strengths of H+R and H+L+R in Table 5 also make the final configuration-selection rule important. In Table 6, multiview regularization changes CVRE L1 only from 0.050 to 0.049; this difference alone does not establish a stable contribution without further evidence under fixed settings. The issue is whether the reported improvements are attributable and reproducible. A practical configuration need not maximize every metric, but its complete performance must be reported together.

2. Essential baselines are missing from the complete comparison, leaving UHD quality gains and the necessity of the proposed design insufficiently established. At 4K/6K, the four evaluated 3DGSSR baselines encounter training OOM, leaving the main tables primarily comparing against 3DGS-bicubic, 3DGS-hr, and Mip-Splatting. This supports feasibility under the specified memory limit, but does not adequately establish quality advantages over feasible alternatives that also use super-resolution priors. The most direct baseline, 3DGS-InvSR, appears in Table 3 for perceptual quality and CVRE but is absent from the complete fidelity and resource comparison. Its 4K MUSIQ on T&T/DB is 75.00 versus 71.01 for the proposed method, while the proposed method has better CVRE. A complete comparison with this strong, simple baseline is therefore necessary to assess the quality, consistency, and cost tradeoff.

The method also uses a Scaffold-GS anchor backbone, without a simplified control using the same representation, SISR prior, and budget. This makes it difficult to separate the effects of the compact backbone, shared RGB/latent geometry, and latent refinement head. Prior latent 3D representations do not automatically negate novelty, but increase the importance of such direct controls. Section 4.1 states both that recommended baseline settings are used and that InvSR is used across 3DGSSR methods; the consequences of replacing their native priors also need clarification.

Finally, captured-scale fidelity is lower than that of 3DGS-bicubic: for example, 2K PSNR on Mip-NeRF 360 is 25.72 versus 27.81 dB. This may be a reasonable perceptual tradeoff, but it needs to be assessed jointly with the above baselines under an explicit input-scale and evaluation protocol. Missing end-to-end runtime, including SISR decoding, further limits the evidence for practical UHD use.

Questions for authors
  1. Clarification and correction of configuration selection. Please list the evaluated configurations and identify the configuration corresponding to each reported metric, including whether test metrics were used for selection. Please report complete quality, CVRE, peak memory, and Gaussian count for one fixed configuration, repeat the RGB-branch ablation under matched settings, and clarify whether the small multiview-regularization gain in Table 6 is stable.

  2. Complete the essential baseline comparisons. Please include 3DGS-InvSR in the full main-table comparison and add a control using the same Scaffold-GS backbone with independent InvSR or a simplified latent branch, to separate backbone gains from those of shared geometry and refinement. For existing 3DGSSR methods, please specify the recommended settings used, whether native priors were replaced, and the impact of replacement. Genuine OOM results under the 24 GB limit should be retained, alongside quality comparisons with strong alternatives that can run within the budget.

  3. Clarification of input and reference scales. Please specify the input resolution, SR factor, reference-image resolution, and prediction-resizing procedure for each target resolution. T&T/DB is described as using 1K inputs, but 3DGS-bicubic has captured-scale PSNR values of 24.62/26.02/26.45 at 1K/2K/4K targets. Which training or evaluation settings change? For scenes with sufficiently high-resolution original captures, is it possible to retain independent test images for reference-based HR evaluation? If not, please explain the limitation imposed by the current data processing.

  4. Report comparable end-to-end costs. Under the fixed configuration above, please report total training time and per-view rendering time including decoding/SISR for the proposed method, 3DGS-hr, and 3DGS-InvSR, with explicit definitions of training and inference memory. The paper already defines OOM as a training failure; the remaining issue is whether UHD inference also retains a resource advantage and what practical cost accompanies the quality gains.

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

This paper presents a method for faithful scene reconstruction and plausible highfrequency synthesis under limited GPU resources. Specifically, it represents observed appearance and super-resolution latent features using two branches with shared Gaussian geometry. A refinement head transforms rendered features into latents for a frozen InvSR decoder, producing high-resolution images without target-resolution Gaussian rasterization. Moreover, a multiview regularizer in latent space is exploited to improve multiview consistency. The experiments validate the effectiveness of the proposed method.

Critical strengths
  1. The paper is generally clear and easy to follow.
  2. The dual-branch design, i.e., one for geometry and low-res details and one for latent features, is reasonable, although not novel enough.
  3. The Latent-Aware Multi-View Regularization module is also sound and helpful for improving multi-view consistency.
Critical weaknesses
  1. The whole pipeline is incremental. Attaching latent features to Gaussians and decoding rendered features builds on established latent-rendering ideas. The dual-branch design is also not sufficiently distinguished from a straightforward extension.
  2. Depth reprojection with depth, direction, and appearance weights is a familiar consistency strategy. It is also not very clear to me what the advantages of applying it in latent space are.
  3. In the ablation study part, it seems that for different ablations, different metrics are utilized. More consistent metric evaluation is expected.
  4. No video supplementary material is provided. This makes it harder to evaluate the multi-view consistency as well as the quality of the rendered results.
Questions for authors

Please refer to the Weaknesses part.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.