A$^2$TG: Adaptive Anisotropic Textured Gaussians for Efficient 3D Scene Representation

Hung-Kuo Chu; Sheng-Chi Hsu; Shih-Hsuan Hung; Ting-Yu Yen

REVIEW 4 major objections 5 minor 1 cited by

Per-Gaussian textures that adapt resolution and aspect ratio match fixed-texture quality while using far less memory.

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 ↗

Adaptive anisotropic texture allocation reduces memory in textured Gaussian splatting while keeping rendering quality competitive.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A useful incremental improvement for textured Gaussian splatting with real memory savings, but the gradient-selection derivation has a gap and the claims need tightening. the 4 major comments →

arxiv 2601.09243 v2 pith:M3ZEA2HS submitted 2026-01-14 cs.CV

A²TG: Adaptive Anisotropic Textured Gaussians for Efficient 3D Scene Representation

Sheng-Chi Hsu , Ting-Yu Yen , Shih-Hsuan Hung , Hung-Kuo Chu This is my paper

classification cs.CV

keywords Gaussian splattingtextured Gaussiansadaptive texture allocationanisotropic texture3D scene representationnovel view synthesismemory efficiencygradient-guided upscaling

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.

The reading

The paper tries to establish that textured Gaussian splatting can be made memory-efficient without sacrificing image quality if each Gaussian carries its own anisotropic texture, sized and shaped according to how much high-frequency detail it actually contributes. It proposes a gradient-guided rule that upscales a texture only when the accumulated positional gradient for that Gaussian passes a threshold, and then doubles the texture along the axis of the splat's anisotropy. Across Mip-NeRF 360, Tanks and Temples, and DeepBlending, the method matches or approaches the rendering fidelity of fixed-texture baselines while using a fraction of the texture memory, and it beats them under a fixed memory budget. A sympathetic reader would care because texture-based Gaussian splatting is a promising route to photorealistic real-time rendering, but its memory cost has been the main obstacle; this work attacks exactly that bottleneck.

Core claim

The central claim is that texture capacity in a 2D Gaussian splatting scene is best spent non-uniformly: each Gaussian gets a rectangular RGBA texture whose width, height, and aspect ratio are decided jointly by the magnitude of its accumulated positional gradient and the ratio of its two semi-axes. Gaussians with large gradient magnitudes (evidence of unresolved high-frequency content) are upscaled, and the upscaling is applied anisotropically when the splat is elongated. The result, the paper reports, is a textured-GS representation with consistently lower memory than fixed 4×4 square textures at matched quality: for example, on DeepBlending at a fixed 500k Gaussian budget, A2TG uses 140 M

What carries the argument

The gradient-guided adaptive texture control: after each optimization step the accumulated positional gradient magnitude ||∇_μ L||_2 (Equations 6–8) is compared against a threshold k_G; Gaussians above it are selected as candidates. Their textures are then upscaled every 500 iterations by doubling T_u and/or T_v depending on the semi-axis ratio s_x/s_y relative to thresholds k_A and k_S. This couples the texture-allocation decision to the actual reconstruction error signal, so texture parameters are spent where the loss is still sensitive to the Gaussian's position.

Load-bearing premise

The entire adaptive allocation rests on the assumption that a Gaussian's accumulated positional gradient magnitude is a reliable proxy for how much texture resolution it needs; if that proxy misidentifies texture-needy primitives, the allocation could waste capacity on well-reconstructed regions and starve detailed ones.

What would settle it

Train A2TG on a scene where a planar, low-frequency region (e.g., a blank wall) is adjacent to a thin, high-frequency anisotropic detail (e.g., a wire fence) viewed at grazing angle. If the wall's Gaussians receive high texture resolution because their positional gradients are inflated by the surrounding detail, while the fence's Gaussians receive low resolution due to occlusion, then the method's memory–quality trade-off should degrade; measure the distribution of texture resolutions and the PSNR in those regions.

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

If this is right

Under a fixed memory budget, A2TG achieves higher PSNR/SSIM and lower LPIPS than fixed-texture baselines on Mip-NeRF 360, Tanks & Temples, and DeepBlending (Table 1).
At a fixed number of Gaussians, A2TG requires roughly 20–30% more memory than plain 2DGS (compared to ~110% for fixed 4×4 textures), with quality close to or better than the fixed-texture methods (Table 2).
The ablation shows that turning off upscaling drops PSNR by ~0.3–0.4 dB, while disabling anisotropic upscaling costs ~0.15–0.2 dB but adds ~10 MB; so most of the gain comes from adaptive allocation, and anisotropy mainly saves memory.
The adaptive scheme naturally leaves most Gaussians at 1×1 textures, so the representation is well-suited to scenes with mixed texture complexity.

Where Pith is reading between the lines

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

The same gradient-driven allocation principle could be applied to other per-primitive attributes (e.g., SH degree, opacity resolution), not just texture dimensions; the paper does not explore this.
Because the gradient signal is computed from the rendering loss, the method should degrade gracefully if the scene is captured at lower resolution: fewer pixels per Gaussian means lower accumulated gradients, so fewer textures upscale — a testable prediction.
The texture-atlas packing in the appendix suggests a path to GPU-friendly compression: rather than storing a full texture per Gaussian, the atlas could be encoded once and shared across splats, which would further cut memory.
The paper's reliance on MCMC densification for the first stage hints that adaptive texture control could be combined with splat-count reduction or quantization methods, but no such combination is demonstrated.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A useful incremental improvement for textured Gaussian splatting with real memory savings, but the gradient-selection derivation has a gap and the claims need tightening.

read the letter

Bottom line: this is a solid, modest efficiency paper for textured Gaussian splatting. The main idea — replace fixed square per-Gaussian textures with per-primitive rectangular textures whose resolution and aspect ratio are grown according to a gradient-based selection rule — is a reasonable extension and the memory savings are real in the reported tables. The paper should be engaged with, but it needs revision before the claims as stated are accepted.

What's new: prior textured-GS methods use a fixed 4×4 texture per Gaussian. The authors instead let each Gaussian have a texture in {1,2,4}×{1,2,4}, upscaling only the primitives whose accumulated positional-gradient magnitude exceeds a threshold and doubling along the axis of larger Gaussian extent. That is a concrete, sensible mechanism and the ablations show it uses less memory than the isotropic variant at nearly the same quality. The CUDA atlas packing for variable-resolution textures is also a useful implementation detail. I credit the fixed-Gaussian-count tables: at 500k Gaussians A2TG is about 140-151MB versus 244MB for the reimplemented Textured Gaussians* with PSNR differences around 0.1-0.2 dB in either direction. That is the honest core of the paper.

Soft spots: the position-gradient derivation in Section 4.2 is not the full gradient. The rendered color depends on μ_i through the texture lookup T(u(x)) because the ray-splat intersection changes when the Gaussian moves. Equations (6)-(8) only propagate through α_i and they drop ∂T/∂u·∂u/∂μ in both the RGB and alpha paths. I don't think this is a fatal error — the selected Gaussian set may still be useful in practice — but as written the 'gradient-guided' claim needs either a full autodiff gradient or an explicit statement that this is an approximation plus some evidence it doesn't change the selection much. The hand-set thresholds kG, kA, kS and the 4×4 cap have no sensitivity analysis, and the memo that they are tuned on the same benchmark is a fair point. The paper also gives no error bars or per-scene variance, and the fixed-memory-budget construction is underspecified. The abstract's phrase 'enhancing image quality' is too strong — Table 2 mostly shows slightly lower PSNR/SSIM than Textured Gaussians* at the same Gaussian count, with memory the clear win. The baseline is an unofficial reimplementation, so the numbers should be taken with some salt. No code is released, which at this stage of the field is a real obstacle.

Who is this for: researchers working on compact Gaussian representations, especially anyone trying to push texture-based splatting toward memory budgets. It's not a paradigm shift; it's a useful incremental contribution.

Recommendation: yes, send it to peer review. The core idea is sound, the empirical evidence is sufficient to justify referee time, and the gradient gap is fixable in revision. I'd ask the authors to release code and to report the full positional gradient comparison.

Referee Report

4 major / 5 minor

Summary. The paper proposes A2TG, a textured Gaussian splatting representation in which each 2D Gaussian is assigned an anisotropic RGBA texture whose resolution and aspect ratio are grown adaptively during training. The growth rule uses the accumulated positional gradient of each Gaussian (Eqs. 6-8) together with the Gaussian's anisotropy (Sec. 4.3) to decide whether to double one or both texture dimensions. The method is evaluated against 2DGS, 2DGS-MCMC, Super Gaussians, BBSplat, and an unofficial 2DGS-based reimplementation of Textured Gaussians on Mip-NeRF 360, Tanks and Temples, and DeepBlending, under both fixed-memory and fixed-Gaussian-count protocols. The main claims are that A2TG achieves comparable or better fidelity than fixed-texture textured Gaussians with substantially lower memory, and that anisotropic allocation is responsible for most of the memory savings. The paper includes ablations, texture-resolution sweeps, a texture-shape distribution analysis, and runtime measurements.

Significance. If the result holds, A2TG provides a useful memory-quality trade-off for textured Gaussian splatting: it concentrates texture parameters on Gaussians that need them, while keeping a compact representation elsewhere. The paper has several concrete strengths: it reports fixed-Gaussian-count tables that make the memory savings explicit (e.g., Table 2, 500k row: 244 MB for Textured Gaussians* vs 140-150 MB for A2TG), it includes ablations that isolate the effect of upscaling and anisotropy (Table 3), it studies the effect of the maximum texture cap (Table 6), and it provides a per-scene distribution of texture shapes (Fig. 4). The central idea is plausible and complementary to existing compression methods. However, the gradient signal that drives the adaptive allocation is not fully derived, and the sensitivity of the method to its hand-set thresholds is not reported. These issues must be addressed before the claimed efficiency advantage can be considered robust.

major comments (4)

[Sec. 4.2, Eqs. (6)-(8)] The positional gradient used for texture selection is not the true gradient of the loss with respect to the Gaussian position. Eq. (6) chains the loss only through ∂α_i/∂μ_i, but the rendered color in Eq. (5) also depends on μ_i through the UV coordinate u(x_j), which enters the texture lookups T_i^RGB(u(x_j)) and T_i^A(u(x_j)) in Eqs. (3)-(4). Once textures are non-constant (i.e., after the first upscaling), ∂c_i(x_j)/∂μ_i contains additional terms involving ∂T_i^RGB/∂u · ∂u/∂μ_i, and Eq. (8) omits the corresponding ∂T_i^A/∂u · ∂u/∂μ_i term. These omitted terms can be large precisely for Gaussians with high-frequency texture content, which are the intended candidates for upscaling. Thus the selection rule is not grounded in the gradient it claims to use. The paper should either derive the full gradient or provide empirical evidence that the omitted terms are negligible, e.g., by compari
[Sec. 5.1 and Sec. 5.3] The adaptive rule depends on four hand-set constants: kG = 0.00002, kA = 4.0, kS = 0.01, and the maximum texture cap {1,2,4}×{1,2,4}, plus the 500-iteration upscaling interval. No sensitivity analysis is reported for kG, kA, or kS, and the interval/cap are only partially varied in Table 6. Because the central claim is that the gradient-guided anisotropic allocation, not these specific thresholds, yields the efficiency gain, the paper should report how memory and quality vary when kG, kA, and kS are perturbed around their chosen values. Without this, it is possible that the reported advantage is a narrow consequence of threshold tuning on the same benchmark suite.
[Sec. 5.1, Table 2] The primary textured baseline, Textured Gaussians*, is an unofficial implementation that replaces 3DGS rasterization with 2DGS. The paper discloses this, but the central quality comparison against this baseline depends on how faithfully it represents the published Textured Gaussians method. The authors should provide a justification that the reimplementation is equivalent apart from the intended rasterization change, or report results obtained with the official implementation. Otherwise the quantitative claims of 'comparable or better quality' are difficult to evaluate.
[Sec. 4.3 and Sec. 5.1] The upscaling schedule is described inconsistently. Sec. 4.3 states that adaptive texture upscaling is 'applied every 500 iterations,' while Sec. 5.1 says A2TG 'applies adaptive texture upscaling at iterations 500 and 1000.' These are different procedures: the former would produce many more upscaling events over 30,000 iterations. The implementation details and the reported results should clarify the actual schedule, and the paper should at least ablate the choice of when upscaling stops, since it directly controls texture capacity.

minor comments (5)

[Abstract and Sec. 5.2] The abstract states that A2TG 'consistently outperforms' fixed-texture Gaussian Splatting methods, but under a fixed number of Gaussians (Table 2) Textured Gaussians* often achieves higher PSNR and SSIM. The claim should be qualified to the fixed-memory regime or to the memory-quality trade-off.
[Table 3] The caption says 'top three results are highlighted in red, orange, and yellow' but there are only three methods in each block; the highlighting is uninformative and should be removed or the caption corrected.
[Sec. 4.3] The scale threshold kS = 0.01 is introduced without defining the units of sx and sy or whether these are normalized scale values. This makes the rule hard to interpret or reproduce.
[Appendix A.2] Table 5 reports FPS/training times for a single scene; please state how many runs and whether these are representative. The main text also lacks per-scene variance or error bars, which would strengthen the quantitative comparisons.
[Throughout] There are several typos and awkward phrases: 'expecially' (Sec. 5.2), 'more then 4 times' (Sec. 5.2), 'efficiency of over method' (Conclusion), and the phrase 'T A_i (u(x))and·T A_i (u(x))' near Eq. (4). These should be cleaned up.

Circularity Check

0 steps flagged

No significant circularity: the adaptive texture upscaling is an optimization heuristic validated by external benchmarks, not a result that reduces to its own inputs.

full rationale

The paper's central mechanism is a training-time adaptive refinement rule: Section 4.2 selects Gaussians for texture upscaling using an accumulated positional-gradient threshold (Eqs. 6-8), and Section 4.3 doubles texture width/height according to semi-axis ratios. The claimed payoff—comparable rendering fidelity with lower memory—is established empirically on external benchmark datasets (Mip-NeRF 360, Tanks & Temples, DeepBlending), with metrics reported in Tables 1-6. The gradient-selection rule is an optimization heuristic analogous to 3DGS densification; final quality emerges from subsequent optimization of the upscaled textures and is not defined by the selection rule. The hand-set thresholds (kA=4.0, kS=0.01, kG=0.00002) and the maximum 4x4 texture cap are hyperparameters; no equation reduces the reported PSNR/SSIM/LPIPS numbers to these constants, and there is no fitted quantity being renamed as a prediction. The paper cites only external prior work, with no load-bearing self-citation or imported uniqueness theorem. The skeptic's concern that Eqs. (6)-(8) omit the ∂T/∂u · ∂u/∂µ terms in the positional gradient is a potential correctness or implementation gap in the stated gradient formula, not a circularity: even if the selection signal is an approximate rather than exact gradient, it does not make the evaluation metric equal to the selection input. No circular step can be exhibited by the paper's own equations, so the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

The central claim rests on several hand-set thresholds (kG, kA, kS), a maximum texture cap, and the upscaling interval. It also assumes that positional gradients and Gaussian semi-axis ratios are good predictors of texture usefulness. No new physical or conceptual entity is introduced.

free parameters (5)

kG = 0.00002
Gradient threshold for selecting Gaussians whose texture resolution should be increased; set by hand in Section 5.1.
kA = 4.0
Anisotropy ratio threshold that decides whether to double only one texture dimension; set by hand in Section 5.1.
kS = 0.01
Small-scale threshold used together with kA to avoid upscaling the longer axis of tiny Gaussians; set by hand in Section 5.1.
maximum texture resolution cap = 4x4
Adaptive upscaling is capped at {1,2,4} x {1,2,4}; this cap is an experimental choice that limits memory and is not derived from a principle.
texture upscaling interval = 500 iterations
Upscaling is applied every 500 iterations; the interval is a hand-set schedule that affects how often gradients are accumulated before resolution changes.

axioms (3)

domain assumption Accumulated positional gradient magnitude identifies texture-needy Gaussians.
Section 4.2 uses ||grad_mu_i L||_2 > kG to select Gaussians for texture upscaling; if this proxy fails, texture allocation is misdirected.
domain assumption Gaussian semi-axis ratio determines the useful texture aspect ratio.
Section 4.3 doubles only the axis with larger scale when the ratio exceeds kA; this assumes rectangular textures aligned with the Gaussian axes are beneficial.
standard math The alpha-compositing gradient equations from 2DGS are correct.
Equations 6-8 derive positional gradients from Eq. 5; the math is standard but is the backbone of the selection signal, so an error here would break the method.

reviewed 2026-08-03 · how reviews work

0 comments

Cite this review

Pith. "Pith review of A$^2$TG: Adaptive Anisotropic Textured Gaussians for Efficient 3D Scene Representation." pith.science (2026). https://pith.science/paper/M3ZEA2HS

@misc{pith2026260109243,
  author       = {Pith},
  title        = {Pith review of: A$^2$TG: Adaptive Anisotropic Textured Gaussians for Efficient 3D Scene Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3ZEA2HS}},
  note         = {Machine review of arXiv:2601.09243}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

Gaussian Splatting has emerged as a powerful representation for high-quality, real-time 3D scene rendering. While recent works extend Gaussians with learnable textures to enrich visual appearance, existing approaches allocate a fixed square texture per primitive, leading to inefficient memory usage and limited adaptability to scene variability. In this paper, we introduce adaptive anisotropic textured Gaussians (A$^2$TG), a novel representation that generalizes textured Gaussians by equipping each primitive with an anisotropic texture. Our method employs a gradient-guided adaptive rule to jointly determine texture resolution and aspect ratio, enabling non-uniform, detail-aware allocation that aligns with the anisotropic nature of Gaussian splats. This design significantly improves texture efficiency, reducing memory consumption while enhancing image quality. Experiments on multiple benchmark datasets demonstrate that A TG consistently outperforms fixed-texture Gaussian Splatting methods, achieving comparable rendering fidelity with substantially lower memory requirements. Project page: http://github.com/Rickyeeeeee/A2TG.

Figures

Figures reproduced from arXiv: 2601.09243 by Hung-Kuo Chu, Sheng-Chi Hsu, Shih-Hsuan Hung, Ting-Yu Yen.

**Figure 1.** Figure 1: Overview of gradient-based adaptive texture control. Given an initial 2DGS model, (a) our system first optimize the parameters of the 2D Gassians and their textures. (b) Next, we compute the positional gradient of the textured 2D Gassians (as depicted in gray blocks) and select the 2D Gassians that need to increase the resolution of the texture to gain more details. (c) Finally, we adaptively upscale the t… view at source ↗

**Figure 2.** Figure 2: Comparison of 2DGS, A2TG and Textured Gaussians* on the DeepBlending datasets. Left: PSNR versus memory size (MB). Right: memory size (MB) versus point count. A2TG achieves higher reconstruction quality under the same memory budget and requires less memory than Textured Gaussians for the same number of Gaussians. (a) GT (b) 2DGS* (c) Textured Gaussians* (d) A2TG [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: Qualitative comparisons. We show the qualitative comparisons of 2DGS* and Textured Gaussians* with A2TG under fixed memory constraint from the Mip-NeRF360 datasets and the DeepBlending datasets. With textures, both Textured Gaussians* and A2TG reconstruct fine scene details, whereas A2TG uses less memory. we set the spherical harmonics color cSHi to zero, leaving appearance determined solely by the learned… view at source ↗

**Figure 4.** Figure 4: The percentage and distribution. This figure shows the percentage and distribution of texture resolution produced by the adaptive texture upscaling on the scene Garden from MipNerf360 dataset. Gaussians highlighted in blue have square texture of 2 × 2 and 4 × 4, and those highlighted in red have non-square texture resolution [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Qualitative visualization of what the adaptive textures learn. Left: full rendering from A 2TG. Middle: rendering without textures (RGB textures set to zero, alpha textures set to one). Right: rendering without SH base color. The comparison, visualized on two scenes shows that textures capture high-frequency residual appearance such as foliage structure and fabric detail, while SH color provides smooth, lo… view at source ↗

**Figure 6.** Figure 6: Comparison of 2DGS, A2TG and Textured Gaussians* on the DeepBlending datasets. Left: PSNR versus memory size (MB). Right: memory size (MB) versus point count. A2TG achieves higher reconstruction quality under the same memory budget and requires less memory than Textured Gaussians for the same number of Gaussians. A.3 ADDITIONAL EXPERIMENTS ON TEXTURE MAP RESOLUTION To further study how texture resolution a… view at source ↗

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Compact Representation of Mipmapped SVBRDFs via Shared Gaussians
cs.GR 2026-07 conditional novelty 6.0

Shared 2D Gaussians across mip levels and SVBRDF maps beat ASTC on quality-per-byte while keeping random-access, non-neural GPU decoding.

Reference graph

Works this paper leans on

8 extracted references · 5 linked inside Pith · cited by 1 Pith paper

[3]

Taming 3DGS: High-quality radiance fields with limited resources

Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3DGS: High-quality radiance fields with limited resources. InSIGGRAPH Asia 2024 Conference Papers, pp. 1–11,

2024
[4]

Hdgs: Textured 2d gaussian splatting for enhanced scene rendering.arXiv preprint arXiv:2412.01823,

Yunzhou Song, Heguang Lin, Jiahui Lei, Lingjie Liu, and Kostas Daniilidis. Hdgs: Textured 2d gaussian splatting for enhanced scene rendering.arXiv preprint arXiv:2412.01823,

Pith/arXiv arXiv
[5]

BillBoard Splatting (BB- Splat): Learnable textured primitives for novel view synthesis.arXiv preprint arXiv:2411.08508,

David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. BillBoard Splatting (BB- Splat): Learnable textured primitives for novel view synthesis.arXiv preprint arXiv:2411.08508,

Pith/arXiv arXiv
[6]

Gaussian billboards: Expressive 2d gaussian splatting with textures.arXiv preprint arXiv:2412.12734,

Sebastian Weiss and Derek Bradley. Gaussian billboards: Expressive 2d gaussian splatting with textures.arXiv preprint arXiv:2412.12734,

Pith/arXiv arXiv
[7]

SuperGaussians: Enhancing gaussian splatting using primitives with spatially varying colors.arXiv preprint arXiv:2411.18966, 2024a

Rui Xu, Wenyue Chen, Jiepeng Wang, Yuan Liu, Peng Wang, Lin Gao, Shiqing Xin, Taku Komura, Xin Li, and Wenping Wang. SuperGaussians: Enhancing gaussian splatting using primitives with spatially varying colors.arXiv preprint arXiv:2411.18966, 2024a. Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, and Song-Hai Zhang. Texture-GS: Disentan- gling the geometry ...

Pith/arXiv arXiv
[8]

During rendering, after computing the UV coordinate of the intersection between a pixel and a Gaussian along the camera ray, we fetch texture values using bilinear interpolation

by augmenting each Gaussian with a non-uniform, variable-sized texture. During rendering, after computing the UV coordinate of the intersection between a pixel and a Gaussian along the camera ray, we fetch texture values using bilinear interpolation. To efficiently support non-uniform texture resolutions ranging from{1,2,4} × {1,2,4}, we pack all per-Gaus...

2024
[2024]

Deep blending for free-viewpoint image-based rendering

Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Bros- tow. Deep blending for free-viewpoint image-based rendering. 37(6):257:1–257:15, 2018a. Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Bros- tow. Deep blending for free-viewpoint image-based rendering.ACM Transactions ...

2024
[2025]

Textured-GS: Gaussian splatting with spatially defined color and opacity.arXiv preprint arXiv:2407.09733,

Zhentao Huang and Minglun Gong. Textured-GS: Gaussian splatting with spatially defined color and opacity.arXiv preprint arXiv:2407.09733,

Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.

[1] [3]

Taming 3DGS: High-quality radiance fields with limited resources

Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3DGS: High-quality radiance fields with limited resources. InSIGGRAPH Asia 2024 Conference Papers, pp. 1–11,

2024

[2] [4]

Hdgs: Textured 2d gaussian splatting for enhanced scene rendering.arXiv preprint arXiv:2412.01823,

Yunzhou Song, Heguang Lin, Jiahui Lei, Lingjie Liu, and Kostas Daniilidis. Hdgs: Textured 2d gaussian splatting for enhanced scene rendering.arXiv preprint arXiv:2412.01823,

Pith/arXiv arXiv

[3] [5]

BillBoard Splatting (BB- Splat): Learnable textured primitives for novel view synthesis.arXiv preprint arXiv:2411.08508,

David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. BillBoard Splatting (BB- Splat): Learnable textured primitives for novel view synthesis.arXiv preprint arXiv:2411.08508,

Pith/arXiv arXiv

[4] [6]

Gaussian billboards: Expressive 2d gaussian splatting with textures.arXiv preprint arXiv:2412.12734,

Sebastian Weiss and Derek Bradley. Gaussian billboards: Expressive 2d gaussian splatting with textures.arXiv preprint arXiv:2412.12734,

Pith/arXiv arXiv

[5] [7]

SuperGaussians: Enhancing gaussian splatting using primitives with spatially varying colors.arXiv preprint arXiv:2411.18966, 2024a

Rui Xu, Wenyue Chen, Jiepeng Wang, Yuan Liu, Peng Wang, Lin Gao, Shiqing Xin, Taku Komura, Xin Li, and Wenping Wang. SuperGaussians: Enhancing gaussian splatting using primitives with spatially varying colors.arXiv preprint arXiv:2411.18966, 2024a. Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, and Song-Hai Zhang. Texture-GS: Disentan- gling the geometry ...

Pith/arXiv arXiv

[6] [8]

During rendering, after computing the UV coordinate of the intersection between a pixel and a Gaussian along the camera ray, we fetch texture values using bilinear interpolation

by augmenting each Gaussian with a non-uniform, variable-sized texture. During rendering, after computing the UV coordinate of the intersection between a pixel and a Gaussian along the camera ray, we fetch texture values using bilinear interpolation. To efficiently support non-uniform texture resolutions ranging from{1,2,4} × {1,2,4}, we pack all per-Gaus...

2024

[7] [2024]

Deep blending for free-viewpoint image-based rendering

Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Bros- tow. Deep blending for free-viewpoint image-based rendering. 37(6):257:1–257:15, 2018a. Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Bros- tow. Deep blending for free-viewpoint image-based rendering.ACM Transactions ...

2024

[8] [2025]

Textured-GS: Gaussian splatting with spatially defined color and opacity.arXiv preprint arXiv:2407.09733,

Zhentao Huang and Minglun Gong. Textured-GS: Gaussian splatting with spatially defined color and opacity.arXiv preprint arXiv:2407.09733,

Pith/arXiv arXiv