Pith. sign in

REVIEW 2 major objections 5 minor 50 references

Pixel-aligned 3D Gaussian rendering can reach high resolution without paying quadratic compute, by subdividing coarse Gaussians into sub-pixel primitives.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 09:17 UTC pith:5EGVQOX2

load-bearing objection Solid 2× result for feed-forward 3DGS, but the ×4 SOTA claim is a zero-shot resolution transfer that needs validation and error bars. the 2 major comments →

arxiv 2607.20813 v1 pith:5EGVQOX2 submitted 2026-07-23 cs.CV

SubSplat: High-Resolution Pixel-aligned 3DGS via Sub-pixel Gaussian Reparameterization

Jiun Lee , Jaekwang Kim , Sangmin Lee This is my paper
classification cs.CV
keywords sub-pixel Gaussian reparameterizationpixel-aligned 3D Gaussian splattinghigh-resolution novel view synthesisgeneralizable neural renderingdeformable attentionopacity redistributionresolution scalabilitymulti-view feature aggregation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SubSplat takes on a trade-off in pixel-aligned 3D Gaussian splatting: feeding the network high-resolution inputs scales cost quadratically, while low-resolution inputs leave too few Gaussians to render sharp views. The paper claims that a Sub-pixel Gaussian Reparameterizer (SPGR) can subdivide each grid-anchored Gaussian into a set of fine-grained primitives, restoring structural density and high-frequency details directly from low-resolution features. On RealEstate10K and ACID, the method reports higher PSNR, SSIM, and LPIPS than pixel-aligned baselines at both 512×512 and 1024×1024 outputs, while cutting latency from 131 ms to 42 ms against the full-resolution baseline. The upshot: output fidelity can be decoupled from backbone computation, making high-resolution, generalizable novel-view synthesis practical at interactive rates.

Core claim

The paper's central claim is that the resolution bottleneck of pixel-aligned Gaussian splatting can be broken by subdividing each coarse primary Gaussian into K sub-pixel primitives (K = (target/input)^2) rather than by raising the input resolution. A lightweight Sub-pixel Gaussian Reparameterizer predicts per-primitive offsets in position, depth, scale, rotation, color, and a footprint-aware opacity allocation, so the total opacity of each primary Gaussian is conserved across its children. Multi-view deformable attention supplies the geometric and appearance context that guides the subdivision. In the reported experiments, this yields state-of-the-art reconstruction quality at 512×512 and 1

What carries the argument

The Sub-pixel Gaussian Reparameterizer (SPGR). It takes a primary Gaussian (position, scale, quaternion, opacity, SH color) plus aggregated multi-view feature descriptors and outputs K refinements: a sub-pixel offset (∆u, ∆v) and a bounded depth residual along the viewing ray, log-scale and axis-angle rotation residuals, a distribution logit for opacity, and a bounded color gain. A footprint-aware softmax renormalizes opacities so that the sum of child opacities equals the parent's; a minimum screen-space area clamp acts as anti-aliasing. This module, plus three-stage deformable attention for cross-view feature aggregation, is what allows the pipeline to render high-resolution views from low

Load-bearing premise

The ×4 results assume the sub-pixel reparameterizer trained on 128×128 inputs transfers to 256×256 inputs without retraining; if the learned offsets are not scale-invariant, those numbers do not reflect the method as trained for that setting.

What would settle it

Train SubSplat with 256×256 inputs and 1024×1024 targets from 1080p data, and compare PSNR/SSIM/LPIPS against the transferred 128×128-trained model on the same test set. If the directly trained model outperforms the transferred one by more than a small margin, or if the transferred model's numbers shift with input resolution while keeping the ×4 ratio, the reported ×4 result depends on an untested scale-invariance assumption rather than on the method itself.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • High-resolution novel-view synthesis from sparse views becomes feasible at interactive rates (24 FPS reported) without per-scene optimization.
  • Doubling output resolution no longer forces a 4× backbone cost; the cost shifts to a lightweight reparameterizer plus more rasterization work.
  • 3D-level densification beats 2D image-space upscaling: the paper reports higher PSNR/SSIM/LPIPS than bilinear or HiT-SR applied to the same 256×256 inputs.
  • The same ×4-trained model scales from 128×128→512×512 to 256×256→1024×1024, indicating resolution-scalable behavior of the reparameterizer.
  • The +5.69 dB PSNR boost at K=4 and +3.16 dB at K=16 from SPGR suggests subdivision, not merely more Gaussians, carries the fidelity gain.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The untested assumption is cross-resolution transfer: the ×4 numbers come from a model trained at 128×128 input and applied to 256×256 input. A fair comparison would train on 256×256→1024×1024 directly; if the gap is large, the reported state-of-the-art at 1024×1024 is partly a property of the training setup, not the method.
  • The fixed subdivision factor K limits use for arbitrary target resolutions; an adaptive, content-aware K (allocating more primitives to high-frequency regions) is a natural next step, and the paper lists this as future work.
  • The opacity-conservation constraint could be relaxed or learned, trading strict photometric conservation for better detail recovery in heavily occluded regions.
  • The reported latency numbers compare 256×256-input SubSplat against 512×512-input baselines; a like-for-like comparison at identical input resolution would isolate the reparameterizer's overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes SubSplat, an extension of pixel-aligned 3D Gaussian splatting that introduces a Sub-pixel Gaussian Reparameterizer (SPGR). SPGR subdivides each grid-anchored primary Gaussian into K sub-pixel primitives, predicted from low-resolution multi-view features via deformable attention, with opacity redistribution and bounded color/geometry modulations. The method is evaluated on RealEstate10K and ACID at 2× (512×512) and 4× (1024×1024) output scales from 256×256 inputs, reporting state-of-the-art PSNR/SSIM/LPIPS and 3× latency reduction relative to full-resolution baselines. The central claim is that high-resolution rendering can be decoupled from backbone input resolution, avoiding quadratic cost while improving fidelity.

Significance. If the reported results hold, SubSplat addresses a real bottleneck in generalizable pixel-aligned 3DGS: the quadratic increase in network cost when raising input resolution. The architectural idea is well motivated, and the paper includes useful ablations (Tables 4, 5, A, B, C) and direct latency/memory measurements on the same GPU (Table 2, Table E). The method is clearly described and the opacity-conservation formulation in Eq. (17) is principled. However, the ×4 results—a key part of the claim of state-of-the-art across multiple scales—rest on an unvalidated resolution transfer, and no error bars are provided anywhere. The central contribution is promising, but the evidence is not yet complete enough for acceptance.

major comments (2)
  1. [§4.1, Table 1, Table 4] The ×4 results (22.65 PSNR on RealEstate10K at 1024×1024 from 256×256 inputs, and the K=16 ablation row in Table 4) are produced by a model trained on 128×128 inputs targeting 512×512 outputs, then directly applied to 256→1024 inference. The paper states this in §4.1 but provides no validation of the transfer: no comparison against a model trained on 256×256 inputs targeting 1024×1024 outputs, no intermediate-scale check, and no error bars. The learned quantities—sub-pixel offsets in Eq. (8)–(9), the bounded depth residual in Eq. (10) with β_z tuned at 128-input scale (Supp. Table A), and the opacity allocation in Eq. (17)—are resolution-dependent in principle. If the transfer is imperfect, the ×4 numbers and the 'across multiple target scales' claim are not representative of the method as trained for that setting. This is load-bearing and must be addressed, either by retraining/validati
  2. [All experimental tables] No error bars, confidence intervals, or multi-seed runs are reported for any quantitative result. This matters for load-bearing comparisons: for example, Table 2 shows SubSplat and HiSplat at identical SSIM (0.850), and Table 3 shows SubSplat and MVSplat+Bilinear at identical latency (0.042 s). Without variance estimates, the claimed superiority in these cases cannot be assessed. The ablation sweeps (Tables A, B, C) also appear to be single runs. Reporting standard deviations across at least 3 seeds would substantially strengthen the empirical claims.
minor comments (5)
  1. [Eq. (9) vs. Supp. Eq. (10)] The position update in the main text, Eq. (9), uses z_sel Δu_k/f_x and z_sel Δv_k/f_y as lateral offsets, while the supplementary version, Eq. (10), defines lateral offsets via b_u and b_v (one-pixel world-space ray bases). These are not obviously equivalent; please clarify and make the two formulations consistent.
  2. [Eq. (10)] The text says 'depth residual is bounded via atanh mapping,' but the formula is Δẑ_k = tanh(Δz_k). This is a tanh mapping, not atanh. Please correct the wording.
  3. [§4.1] The statement 'Due to the scarcity of high-resolution training data (1080p), we instead train a base model on 720p sequences using 128×128 inputs' is confusing: the 1080p split is used for inference, but the model is trained on 720p downsampled to 128×128. Please clarify whether the model sees any 1080p data during training and why the 720p/128→512 configuration is considered a proxy for 256→1024.
  4. [Table 3 / Table F] The latency comparison with image-space upsamplers appears to report total latency including rendering. Since SubSplat renders 4× or 16× more Gaussians than the baseline, it would be helpful to break down network vs. rendering time to make the cost comparison fully transparent.
  5. [General notation] The symbol K is introduced as 'the squared ratio between target and input resolutions' in §3.3, but for the 128→512 model followed by 256→1024 application, K=16 corresponds to a 4× output scale while the training task is also 4×. Please state explicitly that K is defined by the training task, not the inference resolution, to avoid confusion.

Circularity Check

0 steps flagged

No circularity found; the zero-shot ×4 resolution transfer is a generalization concern, not a definitional reduction.

full rationale

The derivation is self-contained: the SPGR outputs (Eq. 8: Δu, Δv, Δz, Δlog s, Δr; Eq. 14: ℓ, γ) are produced by learned heads g_θ and h_φ, supervised by ground-truth high-resolution RGB (Eq. 19) and evaluated on held-out scenes. No reported metric is fed back as a training target or defined in terms of the network's own outputs. Opacity conservation (Eq. 17) and color modulation (Eq. 18) are constraints, not empirical predictions, so their 'conservation' is by design and is not presented as evidence. The only self-citation (DISR-NeRF [17], related work on diffusion-based super-resolution) is not load-bearing. The ×4 numbers are produced by a 128→512-trained model applied to 256→1024 without a control (Sec. 4.1: 'This configuration is then directly applied to 1024×1024 inference from 256×256 inputs'); this is a real external-validity weakness and should be scored as correctness risk, but it is not circular because the network is not fitted to those 1024×1024 test values and no equation reduces the prediction to its input. Hyperparameter sweeps (Supp. Tables A–C) are standard model selection. Hence no significant circularity is present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim is an empirical method improvement. The method relies on a set of hyperparameters tuned by ablation (β_z, λ_s, δ, A_min, K), on the MVSplat backbone producing good primary Gaussians, on the deformable-attention module from PixelGaussian being suitable for multi-view aggregation, and on the assumption that a model trained at one input resolution transfers to another. No new physical entities are introduced.

free parameters (6)
  • K = 4 or 16 (chosen per target scale)
    Number of sub-pixel primitives per primary Gaussian; fixed by the squared ratio of target to input resolution.
  • β_z = 0.25
    Bounds the depth residual via β_z·tanh(Δz); selected by ablation (Table A) on RealEstate10K for best PSNR/SSIM/LPIPS.
  • λ_s = 0.60
    Scale factor for subdivided Gaussians; approximates 1/√K and tuned by ablation (Table B).
  • δ = 0.25
    Color gain bound γ ∈ [1−δ, 1+δ]; chosen by ablation (Table C) to balance color drift and local adaptability.
  • A_min = 0.25(1/K) for K=4; 0.0625(1/K) for K=16
    Minimum screen-space footprint for anti-aliasing, set manually in units of target pixel scale.
  • λ = 0.05
    LPIPS loss weight in the training objective; standard from prior pixel-aligned methods.
axioms (5)
  • standard math A pinhole camera model with known intrinsics/extrinsics is sufficient for projecting Gaussian anchors and computing screen-space footprints.
    Used in Eq. (9), Eq. (15), and the Jacobian in Supplementary A.3.
  • domain assumption The MVSplat backbone produces primary Gaussians accurate enough that subdivision can improve them; SubSplat inherits MVSplat's limitations.
    Stated in Limitations (Supplementary C): 'as our approach is built upon MVSplat, it can be affected by extreme occlusion and large viewpoint variations'.
  • domain assumption The deformable-attention setup from PixelGaussian/GaussianFormer aggregates multi-view features effectively for sub-pixel prediction.
    Adopted in Supplementary A.2; no task-specific analysis is provided for why this particular attention formulation is sufficient.
  • domain assumption A model trained to subdivide 128×128-input anchors for 512×512 output generalizes to 256×256-input anchors for 1024×1024 output without retraining.
    The ×4 experiments rely on this transfer (Section 4.1); no direct validation is provided.
  • domain assumption The opacity of a subdivided Gaussian can be redistributed among primitives by footprint-weighted softmax such that the integrated opacity is conserved and rendering quality is preserved.
    Eq. (16)-(17) define the redistribution; the validity is supported only indirectly by overall benchmark results.

pith-pipeline@v1.3.0-alltime-deepseek · 16762 in / 15997 out tokens · 154257 ms · 2026-08-01T09:17:08.591825+00:00 · methodology

0 comments
read the original abstract

Pixel-aligned Gaussian splatting enables efficient and generalizable novel-view synthesis. However, high-resolution rendering faces a critical trade-off where increasing input resolution improves detail at the expense of quadratically rising network computational cost. Conversely, maintaining low-resolution inputs stabilizes this cost but results in insufficient Gaussian density and artifacts. To address this, we propose SubSplat, which introduces Sub-pixel Gaussian Reparameterizer(SPGR) to subdivide primary Gaussians into fine-grained primitives, restoring structural density directly from low-resolution features. We further enhance the reparameterization quality through feature aggregation, which effectively captures high-frequency details across multiple views. Experiments on RealEstate10K and ACID demonstrate that SubSplat achieves high-fidelity rendering with superior efficiency. Our results validate that the proposed framework successfully resolves the trade-off between reparameterization fidelity and network computational cost inherent in pixel-aligned Gaussian Splatting.

Figures

Figures reproduced from arXiv: 2607.20813 by Jaekwang Kim, Jiun Lee, Sangmin Lee.

Figure 1
Figure 1. Figure 1: Comparison of High-Resolution Rendering Strategies. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the SubSplat Architecture. SubSplat aggregates multi-view features via Deformable Attention to extract rich geometry and appearance descriptors. These descriptors guide the Sub-pixel Gaussian Reparameterizer(SPGR) to subdivide Gaussians into fine-grained primitives, restoring high-resolution details while incurring only a marginal increase in total inference time. attention module [9, 14] perfo… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results at 512 × 512 screen space. While pixel-aligned base￾lines suffer from geometric instability and blurring when rendering beyond their 256 × 256 input grid, SubSplat(ours) recovers sharp, structurally consistent details on RealEstate10K [47] and ACID [21] through sub-pixel Gaussian reparameterization. DepthSplat [34]. To evaluate the benefits of 3D-level densification, we also com￾pare ag… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results at 1024 × 1024 screen space. Even with a 256 × 256 input, SubSplat(ours) maintains structural consistency and sharp details on RealEstate10K [47] and ACID [21]. In contrast, pixel-aligned baselines exhibit blur￾ring or severe halos when rendering at 4× resolution [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Comparison with Image-Space Upsamplers. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study of SubSplat modules across subdivision scales. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 8 linked inside Pith

  1. [1]

    In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)

    Barron, J.T., Mildenhall, B., Srinivasan, P.P., Tancik, M., Hedman, P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: IEEE/CVF International Conference on Computer Vision (ICCV) (2021)

    2021

  2. [2]

    In: 11th Pacific Conference onComputer Graphics and Applications, 2003

    Botsch, M., Kobbelt, L.: High-quality point-based rendering on modern gpus. In: 11th Pacific Conference onComputer Graphics and Applications, 2003. Proceed- ings. pp. 335–343. IEEE (2003)

    2003

  3. [3]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Charatan,D.,Li,S.L.,Tagliasacchi,A.,Sitzmann,V.:pixelsplat:3dgaussiansplats from image pairs for scalable generalizable 3d reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19457– 19467 (2024)

    2024

  4. [4]

    In: European conference on computer vision

    Chen, A., Xu, H., Esposito, S., Tang, S., Geiger, A.: Lara: Efficient large-baseline radiance fields. In: European conference on computer vision. pp. 338–355. Springer (2024)

    2024

  5. [5]

    In: European conference on computer vision

    Chen, Y., Xu, H., Zheng, C., Zhuang, B., Pollefeys, M., Geiger, A., Cham, T.J., Cai, J.: Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In: European conference on computer vision. pp. 370–386. Springer (2024) 16 Jiun Lee et al

    2024

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, Z., Funkhouser, T., Hedman, P., Tagliasacchi, A.: Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16569–16578 (2023)

    2023

  7. [7]

    In: Proceed- ings CVPR IEEE computer society conference on computer vision and pattern recognition

    Collins, R.T.: A space-sweep approach to true multi-image matching. In: Proceed- ings CVPR IEEE computer society conference on computer vision and pattern recognition. pp. 358–363. Ieee (1996)

    1996

  8. [8]

    In: European conference on computer vision

    Fang,G.,Wang,B.:Mini-splatting:Representingsceneswithaconstrainednumber of gaussians. In: European conference on computer vision. pp. 165–181. Springer (2024)

    2024

  9. [9]

    arXiv preprint arXiv:2410.18979 (2024)

    Fei, X., Zheng, W., Duan, Y., Zhan, W., Tomizuka, M., Keutzer, K., Lu, J.: Pix- elgaussian: Generalizable 3d gaussian reconstruction from arbitrary views. arXiv preprint arXiv:2410.18979 (2024)

    Pith/arXiv arXiv 2024

  10. [10]

    arXiv preprint arXiv:2404.10318 (2024)

    Feng, X., He, Y., Wang, Y., Yang, Y., Li, W., Chen, Y., Kuang, Z., Fan, J., Jun, Y., et al.: Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318 (2024)

    arXiv 2024

  11. [11]

    Cambridge university press (2003)

    Hartley, R., Zisserman, A.: Multiple view geometry in computer vision. Cambridge university press (2003)

    2003

  12. [12]

    In: ACM SIGGRAPH 2024 conference papers

    Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S.: 2d gaussian splatting for geo- metrically accurate radiance fields. In: ACM SIGGRAPH 2024 conference papers. pp. 1–11 (2024)

    2024

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Huang, X., Li, W., Hu, J., Chen, H., Wang, Y.: Refsr-nerf: Towards high fidelity and super resolution view synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8244–8253 (2023)

    2023

  14. [14]

    In: European Conference on Computer Vision

    Huang, Y., Zheng, W., Zhang, Y., Zhou, J., Lu, J.: Gaussianformer: Scene as gaus- sians for vision-based 3d semantic occupancy prediction. In: European Conference on Computer Vision. pp. 376–393. Springer (2024)

    2024

  15. [15]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023)

    2023

  16. [16]

    arXiv preprint arXiv:2407.08447 (2024)

    Kulhanek, J., Peng, S., Kukelova, Z., Pollefeys, M., Sattler, T.: Wildgaussians: 3d gaussian splatting in the wild. arXiv preprint arXiv:2407.08447 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lee, J.L., Li, C., Lee, G.H.: Disr-nerf: Diffusion-guided view-consistent super- resolution nerf. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20561–20570 (2024)

    2024

  18. [18]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1833–1844 (2021)

    2021

  19. [19]

    In: European conference on computer vision

    Liang, Z., Zhang, Q., Hu, W., Zhu, L., Feng, Y., Jia, K.: Analytic-splatting: Anti- aliased 3d gaussian splatting via analytic integration. In: European conference on computer vision. pp. 281–297. Springer (2024)

    2024

  20. [20]

    In: Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Lin, C.Y., Fu, Q., Merth, T., Yang, K., Ranjan, A.: Fastsr-nerf: Improving nerf efficiency on consumer devices with a simple super-resolution pipeline. In: Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6036–6045 (2024)

    2024

  21. [21]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Liu, A., Tucker, R., Jampani, V., Makadia, A., Snavely, N., Kanazawa, A.: Infi- nite nature: Perpetual view generation of natural scenes from a single image. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14458–14467 (2021)

    2021

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Struc- tured 3d gaussians for view-adaptive rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20654–20664 (2024) SubSplat 17

    2024

  23. [23]

    Commu- nications of the ACM65(1), 99–106 (2021)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Commu- nications of the ACM65(1), 99–106 (2021)

    2021

  24. [24]

    ACM transactions on graphics (TOG)41(4), 1–15 (2022)

    Müller,T.,Evans,A.,Schied,C.,Keller,A.:Instantneuralgraphicsprimitiveswith a multiresolution hash encoding. ACM transactions on graphics (TOG)41(4), 1–15 (2022)

    2022

  25. [25]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Nam, S., Sun, X., Kang, G., Lee, Y., Oh, S., Park, E.: Generative densification: Learning to densify gaussians for high-fidelity generalizable 3d reconstruction. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 26683–26693 (2025)

    2025

  26. [26]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rematas, K., Liu, A., Srinivasan, P.P., Barron, J.T., Tagliasacchi, A., Funkhouser, T., Ferrari, V.: Urban radiance fields. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12932–12942 (2022)

    2022

  27. [27]

    In: European Conference on Computer Vision

    Rota Bulò, S., Porzi, L., Kontschieder, P.: Revising densification in gaussian splat- ting. In: European Conference on Computer Vision. pp. 347–362. Springer (2024)

    2024

  28. [28]

    In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques

    Rusinkiewicz, S., Levoy, M.: Qsplat: A multiresolution point rendering system for large meshes. In: Proceedings of the 27th annual conference on Computer graphics and interactive techniques. pp. 343–352 (2000)

    2000

  29. [29]

    arXiv preprint arXiv:1711.02508 (2017)

    Sola, J.: Quaternion kinematics for the error-state kalman filter. arXiv preprint arXiv:1711.02508 (2017)

    Pith/arXiv arXiv 2017

  30. [30]

    arXiv preprint arXiv:2410.06245 (2024)

    Tang, S., Ye, W., Ye, P., Lin, W., Zhou, Y., Chen, T., Ouyang, W.: Hisplat: Hier- archical 3d gaussian splatting for generalizable sparse-view reconstruction. arXiv preprint arXiv:2410.06245 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

    In: Proceedings of the Computer Vision and Pattern Recog- nition Conference

    Wan, Y., Shao, M., Cheng, Y., Zuo, W.: S2gaussian: Sparse-view super-resolution 3d gaussian splatting. In: Proceedings of the Computer Vision and Pattern Recog- nition Conference. pp. 711–721 (2025)

    2025

  32. [32]

    In: Proceedings of the 30th ACM International Conference on Multimedia

    Wang, C., Wu, X., Guo, Y.C., Zhang, S.H., Tai, Y.W., Hu, S.M.: Nerf-sr: High quality neural radiance fields using supersampling. In: Proceedings of the 30th ACM International Conference on Multimedia. pp. 6445–6454 (2022)

    2022

  33. [33]

    arXiv preprint arXiv:2410.02571 (2024)

    Xie, S., Wang, Z., Wang, X., Zhu, Y., Pan, C., Dong, X.: Supergs: Super-resolution 3d gaussian splatting enhanced by variational residual features and uncertainty- augmented learning. arXiv preprint arXiv:2410.02571 (2024)

    Pith/arXiv arXiv 2024

  34. [34]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xu, H., Peng, S., Wang, F., Blum, H., Barath, D., Geiger, A., Pollefeys, M.: Depth- splat: Connecting gaussian splatting and depth. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16453–16463 (2025)

    2025

  35. [35]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45(11), 13941–13958 (2023)

    Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Yu, F., Tao, D., Geiger, A.: Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence45(11), 13941–13958 (2023)

    2023

  36. [36]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yan, Z., Low, W.F., Chen, Y., Lee, G.H.: Multi-scale 3d gaussian splatting for anti-aliased rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20923–20931 (2024)

    2024

  37. [37]

    In: Proceedings of the European conference on computer vision (ECCV)

    Yao, Y., Luo, Z., Li, S., Fang, T., Quan, L.: Mvsnet: Depth inference for unstruc- tured multi-view stereo. In: Proceedings of the European conference on computer vision (ECCV). pp. 767–783 (2018)

    2018

  38. [38]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yoon, Y., Yoon, K.J.: Cross-guided optimization of radiance fields with multi-view image super-resolution for high-resolution novel view synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12428– 12438 (2023)

    2023

  39. [39]

    Advances in Neu- ral Information Processing Systems38, 9767–9792 (2026) 18 Jiun Lee et al

    Younes, M., Boukhayma, A.: Anti-aliased 2d gaussian splatting. Advances in Neu- ral Information Processing Systems38, 9767–9792 (2026) 18 Jiun Lee et al

    2026

  40. [40]

    arXiv preprint arXiv:2406.10111 (2024)

    Yu, X., Zhu, H., He, T., Chen, Z.: Gaussiansr: 3d gaussian super-resolution with 2d diffusion priors. arXiv preprint arXiv:2406.10111 (2024)

    Pith/arXiv arXiv 2024

  41. [41]

    In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition

    Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition. pp. 19447–19456 (2024)

    2024

  42. [42]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Zhang, C., Zou, Y., Li, Z., Yi, M., Wang, H.: Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 9869–9877 (2025)

    2025

  43. [43]

    In: European Conference on Computer Vision

    Zhang, K., Bi, S., Tan, H., Xiangli, Y., Zhao, N., Sunkavalli, K., Xu, Z.: Gs-lrm: Large reconstruction model for 3d gaussian splatting. In: European Conference on Computer Vision. pp. 1–19. Springer (2024)

    2024

  44. [44]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018)

    2018

  45. [45]

    In: European conference on computer vision

    Zhang, X., Zhang, Y., Yu, F.: Hit-sr: Hierarchical transformer for efficient im- age super-resolution. In: European conference on computer vision. pp. 483–500. Springer (2024)

    2024

  46. [46]

    In: European Conference on Computer Vision

    Zhang, Z., Hu, W., Lao, Y., He, T., Zhao, H.: Pixel-gs: Density control with pixel- aware gradient for 3d gaussian splatting. In: European Conference on Computer Vision. pp. 326–342. Springer (2024)

    2024

  47. [47]

    Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learn- ingviewsynthesisusingmultiplaneimages.arXivpreprintarXiv:1805.09817(2018)

    Pith/arXiv arXiv 2018

  48. [48]

    arXiv preprint arXiv:2010.04159 (2020)

    Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020)

    Pith/arXiv arXiv 2010

  49. [49]

    In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques

    Zwicker, M., Pfister, H., Van Baar, J., Gross, M.: Surface splatting. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques. pp. 371–378 (2001)

    2001

  50. [50]

    Zwicker, M., Pfister, H., Van Baar, J.,Gross, M.: Ewasplatting. IEEE Transactions on Visualization & Computer Graphics8(03), 223–238 (2002) SubSplat: High-Resolution Pixel-aligned 3DGS via Sub-pixel Gaussian Reparameterization Supplementary Material We provide additional technical details and extended results that comple- ment the main paper. Appendix A d...

    2002