acceptodds
Under review as a conference paper at ICLR 2027

LegendreImage: Generalized Fourier Expansion Splatting for Image Representation

Abstract

Existing 2D Gaussian image representations mainly scale their representations through the number and spatial allocation of primitives, while the function space represented by each individual primitive remains comparatively limited. We revisit primitive representation and investigate whether richer local function spaces can provide an additional dimension of representation design. We introduce Generalized Fourier Expansion Splatting (GFES), which views the signal carried by each primitive as a function over an intrinsic domain and parameterizes it through a truncated generalized Fourier expansion. This formulation separates the representation of the primitive-carried signal from its spatial localization: a spatial weighting function determines where the primitive contributes, while the function expansion determines how its signal varies over the corresponding domain. The basis is chosen according to the domain, measure, and function space of the represented signal. Following this principle, we instantiate GFES as LegendreImage by combining compact raised-cosine localization with a two-dimensional Legendre polynomial expansion to represent spatially varying color within individual primitives. Adaptive primitive allocation yields LegendreImage+. On DIV2K, LegendreImage+ surpasses GaussianImage++ from the 2.4M-parameter regime onward and reaches 51.83 dB at a 4M trainable-parameter budget, outperforming GaussianImage++ by 3.77 dB. These results show that the per-primitive function space is an important dimension of representation design, complementary to scaling through the number and spatial allocation of primitives.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Representation, 2D Gaussian Splatting, Generalized Fourier Expansion, Orthogonal Basis Functions, Legendre Polynomials, Splatting-based Representation

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_kBh9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript introduces Generalized Fourier Expansion Splatting (GFES), a signal representation method that treats the signal carried by each primitive as a function over an intrinsic domain and represents it using a truncated generalized Fourier expansion. This approach separates the primitive signal from its spatial localization. Using this formulation, the authors propose "LegendreImage", which uses a two-dimensional Legendre polynomial expansion with a compact raised cosine window and, based on this, they next develop "LegendreImage+" which includes adaptive primitive allocation. Numerical experiments on the DIV2K, Kodak, and DIV8K datasets show competitive reconstruction performance at moderate and high parameter budgets.

Critical strengths

The numerical experiments include several datasets and an extensive set of parameter budgets and ablations. The numerical results show that the method is very competitive against state-of-the-art, achieving higher PSNR than GaussianImage++ at moderate and high parameter budgets on DIV2K and Kodak.

Critical weaknesses

The conceptual novelty of the proposed method is rather limited. The idea of representing the signal carried by each primitive as a function has been exploited already, most notably by 3D Gaussian Splatting [1] and 2D Gaussian-Hermite Splatting [2]. Here, the novelty appears to be the selection of basis functions. However, the main LegendreImage implementation only uses first-order Legendre polynomials, so its local function space is simply span{1, u, v}, i.e., a standard affine model. The adaptive allocation strategy is conceptually closely related to GaussianImage++ [3], which already progressively inserts new primitives at pixel level with large reconstruction distortion. Thus, the underlying idea of allocating primitives according to reconstruction error is not new. Finally, Appendix A only provides a one-dimensional scalar result, whereas the proposed method is two-dimensional. This result is a standard local polynomial approximation property associated with the polynomial space PK, but does not imply any specific advantage of the Legendre basis.

Minor issue • Page 5, Lines 224: In Eq. (6), mn ̸= 0 excludes the first-order terms (1, 0) and (0, 1). I believe it should be (m, n) ̸= (0, 0).

[1] Kerbl, Bernhard, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. ``3d gaussian splatting for real-time radiance field rendering.'' ACM Trans. Graph. 42, no. 4 (2023): 139-1.

[2] Yu, Ruihan, Tianyu Huang, Jingwang Ling, and Feng Xu. "2dgh: 2d gaussian-hermite splatting for high-quality rendering and better geometry features." IEEE Transactions on Visualization and Computer Graphics (2025).

[3] Li, Tiantian, Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Jun Zhang, and Yan Wang. ``GaussianImage++: Boosted Image Representation and Compression with 2D Gaussian Splatting.'' In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 8, pp. 6442-6449. 2026.

Questions for authors

The presentation does not make a convincing case of the potential advantages of using Legendre polynomials. Since P0(x) = 1 and P1(x) = x, the first-order Legendre basis used in the main model (K = 1) is exactly the standard affine basis {1, u, v}. Moreover, increasing the order from K = 1 to K = 2 reduces PSNR under the same parameter budget (Table 2). Thus, what practical advantage does the Legendre formulation provide over a standard local affine model?

4 comments
AuthorsOpenReview ↗

Acknowledgments

We sincerely thank the reviewer for the detailed assessment of our work. We particularly appreciate that your assessment considered the breadth of the experiments and the reconstruction results while also examining the relationship between the mathematical formulation, the research contribution, and the practical benefits. Your questions about the first-order affine form and its practical advantages helped us focus more clearly on what this work seeks to reveal and how that understanding leads to concrete representations.

Your comments prompted us to present the connections across methods revealed by GFES more explicitly, explain how this unified understanding relates to the construction of LegendreImage, and distinguish the gains in expressiveness at a fixed primitive count from the gains in efficiency under a fixed parameter budget. We are very grateful for this attention to the relationship between research claims and experimental evidence. It helped us describe the contribution more precisely and make clearer which conclusions are supported by each set of experiments.

We also greatly appreciate your careful examination of the theoretical analysis and mathematical notation. These specific observations complement the broader conceptual questions and helped us improve the coherence of the manuscript from the research motivation through the mathematical formulation to the interpretation of the experiments. We sincerely thank you for the time and effort devoted to this assessment. Below, we respond to each comment and describe the corresponding revisions.

AuthorsOpenReview ↗

Q1: On the Choice of First-Order Legendre Expansion

For scalar local signals, the Legendre expansion with K=1K=1 spans the same function space as a standard local affine model. The Legendre formulation provides a local polynomial hierarchy that includes this affine form and can be extended to higher orders, making truncation order a design variable for adjusting per-primitive expressiveness and parameter cost. Our use of K=1K=1 in the main experiments is an empirical choice within this representation family under the given parameter budget.

Regarding the “practical advantage,” if a task requires keeping the primitive count fixed, increasing the truncation order can enhance per-primitive expressiveness. Original Appendix C / Table 4 demonstrates this benefit with fixed-allocation LegendreImage, using 100,000 primitives, independent RGB coefficients, and no adaptive allocation:

Legendre Truncation OrderParams / PrimitivePrimitivesTrainable-parameter budgetPSNR (dB)Data Source
K=0K=05100,000500,000=5×100,000500{,}000=5\times100{,}00029.24Table 4
K=1K=111100,0001,100,000=11×100,0001{,}100{,}000=11\times100{,}00032.53Table 4
K=2K=220100,0002,000,000=20×100,0002{,}000{,}000=20\times100{,}00035.46Table 4
K=3K=332100,0003,200,000=32×100,0003{,}200{,}000=32\times100{,}00038.29Table 4
K=4K=447100,0004,700,000=47×100,0004{,}700{,}000=47\times100{,}00041.07Table 4
K=5K=565100,0006,500,000=65×100,0006{,}500{,}000=65\times100{,}00043.45Table 4
K=6K=686100,0008,600,000=86×100,0008{,}600{,}000=86\times100{,}00045.67Table 4

In the main experiments, we compare reconstruction quality under the same trainable-parameter budget to evaluate parameter efficiency. Sec. 5.2 / Table 2 of the original manuscript uses LegendreImage+ with adaptive allocation. Nonzero-order expansion coefficients are shared across RGB channels, while the constant terms remain channel-specific. We restate the original Table 2 results below, all under a 4M trainable-parameter budget:

Legendre Truncation OrderParams / PrimitivePrimitivesTrainable-parameter budgetPSNR (dB)Data Source
K=0K=05800,000=⌊4,000,000/5⌋800{,}000=\lfloor4{,}000{,}000/5\rfloor4,000,00047.06Table 2
K=1K=17571,428=⌊4,000,000/7⌋571{,}428=\lfloor4{,}000{,}000/7\rfloor4,000,00051.83Table 2
K=2K=210400,000=⌊4,000,000/10⌋400{,}000=\lfloor4{,}000{,}000/10\rfloor4,000,00051.21Table 2

Increasing the truncation order increases the parameter count per primitive while reducing the number of primitives available under the same budget. In Table 2, the first-order configuration achieves the highest reconstruction quality among the evaluated orders, reflecting the budget trade-off between per-primitive expressiveness and the number of primitives. Therefore, our choice of K=1K=1 is motivated by parameter efficiency and does not represent an upper limit on the per-primitive expressiveness of LegendreImage.

GFES also places local affine models within a unified approximation perspective. For sufficiently smooth scalar signals, Appendix A gives an O(h2)O(h^2) first-order projection error on the normalized local domain, explaining the effectiveness of local affine approximation at small spatial scales. Conversely, the affine form alone is insufficient to provide GFES’s unified interpretation and construction principles across different signal domains, measures, basis functions, and expansion orders. We have added a discussion of this relationship in Sec. 6.3 of the revised manuscript.

AuthorsOpenReview ↗

W: Novelty of GFES

As summarized in the final paragraph of Sec. 1, our contributions span three levels:

  1. Conceptual level: We introduce GFES in Sec. 1, revealing the shared mathematical structure of different primitive representations.
  2. Methodological level: In Sec. 3, we instantiate GFES as two image representation methods, namely LegendreImage with fixed allocation and LegendreImage+ with adaptive allocation.
  3. Experimental level: Main experiments and ablations demonstrate the reconstruction performance, parameter efficiency, and design trade-offs of both methods in Tables 2, 4, 6, and 7.

Specific configurations, allocation strategies, and optimization choices implement and enhance these representations and are not claimed as separate core innovations.

Regarding the reviewer's statement that "the novelty appears to be the selection of basis functions," GFES reveals the intrinsic mathematical connections among existing primitive representations and uses this understanding to guide the construction of new representations. To address the primitive design question posed at the end of the first paragraph of Sec. 1, we inductively identify the common form of signal expansion combined with spatial localization in existing methods such as 3DGS [1] and 2DGH [2], and describe it using classical generalized Fourier expansion. Appendix F / Table 9 of the revised manuscript gives the concrete correspondences:

MethodSignal CoordinatesSignal RepresentationLocalization
GI [4] / GI++ [3] / LIG [5]Local spaceConstant RGBGaussian function
3DGS [1]Viewing directionSpherical harmonicsGaussian function
2DGH† [2]Local spaceHermite polynomialsGaussian function
WIPES (2D) [6]Modulation phaseConstant and cosineGaussian function
LegendreImage / LegendreImage+ (Ours)Local squareLegendre polynomialsCompact raised-cosine window

GI denotes GaussianImage. † The Hermite expansion in 2DGH describes the shape field before its nonlinear activation.

3DGS uses spherical harmonics on the directional domain, while the cosine modulation in WIPES can be written as a finite Fourier representation on the phase domain. Both use Gaussian functions for spatial localization. The signal and localization domains can differ, yet these representations share the same structure.

This shared structure also provides a mathematical starting point for constructing new representations. Classical expansions such as Chebyshev, Jacobi, Zernike, and Fourier–Bessel provide signal representations associated with their respective domains and measures. Combining them with spatial localization expands the range of potential GFES instances beyond Table 9.

From this understanding, we deductively construct two image representation methods: LegendreImage with fixed allocation and LegendreImage+ with adaptive allocation. Both use a Legendre expansion to represent color variation within a bounded local domain and a compact window for spatial localization, as detailed in Sec. 3. The choice of specific basis functions is a result of this construction.

Original Table 4 shows reconstruction gains from higher-order expansion at a fixed primitive count, demonstrating per-primitive expressiveness. In revised Table 2, fixed-allocation LegendreImage improves from 42.38 dB at zeroth order to 48.29 dB at first order under the same 4M trainable-parameter budget, demonstrating the parameter efficiency of local expansion under this budget.

AuthorsOpenReview ↗

W: Brief Clarifications Regarding Other Comments

Adaptive primitive allocation. Residual-driven allocation has been investigated in prior work, such as GaussianImage++ [3]. We adopt it as a technical enhancement and do not claim it as a main contribution. Fig. 4 and Table 7 show that, under trainable-parameter budgets of 4M, 5.6M, and 7.2M on DIV2K, LegendreImage still achieves 48.29, 53.65, and 56.89 dB, respectively, after removing densification and relocation, exceeding the 48.06, 50.79, and 52.40 dB achieved by GaussianImage++, which already uses adaptive allocation. These results show that the reconstruction advantages at these budgets do not arise entirely from adaptive allocation.

Regarding Appendix A. The O(hK+1)O(h^{K+1}) bound is a standard local polynomial approximation result for smooth scalar functions. It explains how local scale and truncation order jointly affect approximation accuracy, informing local expansion design within GFES without establishing an advantage specific to the Legendre basis. The current analysis is one-dimensional. Under appropriate smoothness and domain regularity assumptions, it can be extended via the Bramble–Hilbert lemma [7] to two-dimensional polynomial spaces of total degree at most KK.

Regarding Eq. (6). We thank the reviewer for pointing out this typographical error. The summation condition should be (m,n)≠(0,0)(m,n)\neq(0,0), excluding the zeroth-order term while retaining all nonzero-order terms. We have corrected this in the revised manuscript.

References

[1] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), Article 139, 2023.

[2] Ruihan Yu, Tianyu Huang, Jingwang Ling, and Feng Xu. 2DGH: 2D Gaussian-Hermite Splatting for High-Quality Rendering and Better Geometry Features. IEEE Transactions on Visualization and Computer Graphics, 2025.

[3] Tiantian Li, Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Jun Zhang, and Yan Wang. GaussianImage++: Boosted Image Representation and Compression with 2D Gaussian Splatting. Proceedings of the AAAI Conference on Artificial Intelligence, 40(8):6442–6449, 2026.

[4] Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Yan Wang, Hongwei Qin, Guo Lu, Jing Geng, and Jun Zhang. GaussianImage: 1000 FPS Image Representation and Compression by 2D Gaussian Splatting. European Conference on Computer Vision, pp. 327–345, 2024.

[5] Lingting Zhu, Guying Lin, Jinnan Chen, Xinjie Zhang, Zhenchao Jin, Zhao Wang, and Lequan Yu. Large Images Are Gaussians: High-Quality Large Image Representation with Levels of 2D Gaussian Splatting. Proceedings of the AAAI Conference on Artificial Intelligence, 39(10):10977–10985, 2025.

[6] Wenhao Zhang, Hao Zhu, Delong Wu, Di Kang, Linchao Bao, Xun Cao, and Zhan Ma. WIPES: Wavelet-based Visual Primitives. Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 27338–27347, 2025.

[7] James H. Bramble and Stephen R. Hilbert. Estimation of Linear Functionals on Sobolev Spaces with Application to Fourier Transforms and Spline Interpolation. SIAM Journal on Numerical Analysis, 7(1):112–124, 1970.

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

The paper studies how local function expansions can increase the expressive capacity of individual primitives in 2D image representation. The paper introduces GFES, and instantiate it as LegendreImage, which enables individual primitives to represent spatially varying local signals within explicitly bounded support. The experiments examine the trade-off between local expressiveness and primitive count and report reconstruction gains at higher parameter budgets.

Critical strengths
  1. The paper provides GFES, a useful formulation for separating local signal expansion from spatial weighting and the algorithm LegendreImage for better single-primitive improvement
  2. The paper provides detailed ablations on high orders, RGB coefficient sharing, and adaptive allocation. The paper also discusses the method's limitations at low parameter budgets.
Critical weaknesses
  1. The idea is good, but I still have concerns on the theory and the novelty: GFES builds on classical localized approximation: the Shepard interpolation uses spatial weighting, while the moving least squares uses locally weighted polynomial fitting. Although the paper acknowledges these connections, you should still distinguish the novelty. Moreover, the first-order Legendre basis spans the same affine space {1,u,v} as Hermite and ordinary linear functions. While the experiments are not enough to establish the advantages directly from Legendre. There are also two counterexamples reveal a gap between the scalar theory and the shared-RGB implementation: (A) For Yp=121+tpdY_p=\frac12\mathbf1+t_p d, where d=(1,−1,0)⊤d=(1,-1,0)^\top, the shared-RGB model Y^=AC+(Bb)1⊤\hat Y=AC+(Bb)\mathbf1^\top satisfies $ \min_{C,b}|Y-\hat Y|_F^2=\min_C|Y-AC|_F^2. $ Which means in this case shared higher-order terms provide no improvement in optimal MSE under fixed primitive geometry. (B) For the smooth signal F(hu)=121+κhudF(hu)=\frac12\mathbf1+\kappa hu d, the best single-primitive shared-slope approximation has L2L^2 error 2∣κ∣h/3=Θ(h)2|\kappa|h/\sqrt3=\Theta(h), while independent RGB slopes achieve zero error. Therefore, Appendix A's scalar O(h2)O(h^2) bound does not directly apply to the shared-RGB local model, limiting its use in explaining the reported reconstruction gains.
  2. Experimental concerns: I found insufficient attribution of the main improvements and a potential inconsistency in adaptive allocation. Eqs. (26)–(27) and Table 3 imply that Kodak budgets ≥480k use (R=2) for all newly inserted primitives. This allows at most nine valid pixels per candidate (a 5/4 checkerboard split), below the requirement in Appendix B.1.
Questions for authors
  1. Could you please isolate the contribution of the local color expansion from compact localization, normalized aggregation, and adaptive allocation with further experiments. Moreover, does the gain-based selector described in Section 3.2 and Appendix B.1 accept any candidates on Kodak at budgets of 480k or higher? These evidences would strongly affect my decision.
  2. What causes the sharp PSNR drop of GaussianImage at high budgets in Table 7, for example the Cholesky variant is falling from 40.66 dB at 4M parameters to 29.74 dB at 7.2M.
  3. please double check the boundary of the theory.
6 comments
AuthorsOpenReview ↗

Acknowledgments

We sincerely thank the reviewer for the careful reading of our manuscript and the close examination of its mathematical formulation and implementation conditions. The concrete counterexamples you provided, together with your analysis of candidate-validity conditions and questions about the selector's actual operation, were particularly helpful in improving the manuscript. We greatly appreciate your willingness to examine these details and consider the theoretical assumptions, parameterization choices, and algorithmic behavior together.

These questions gave us clear starting points for further clarifying the method. They helped us state the scope of the scalar analysis more precisely, distinguish the parameter efficiency of the shared-RGB configuration from the expressiveness of the representation family, and use controlled experiments and execution logs to explain the respective roles of local color expansion, adaptive allocation, and its internal mechanisms. Such specific feedback allowed us to address each concern with relevant analysis and evidence, while making the revised explanations easier for readers to understand and verify.

We greatly value this exchange. Your careful attention to the connections between the mathematical reasoning and the implementation has helped us improve the precision of the discussion and the relevance of the experimental evidence to the questions being examined. Below, we respond to each point and explain the corresponding revisions.

AuthorsOpenReview ↗

W1: Novelty of GFES and LegendreImage

Regarding the reviewer's concerns about "the theory and the novelty," our contributions span three levels, as summarized in the final paragraph of Sec. 1:

  1. Conceptual level: We introduce GFES in Sec. 1, revealing the shared mathematical structure of different primitive representations.
  2. Methodological level: In Sec. 3, we instantiate GFES as two image representation methods, namely LegendreImage with fixed allocation and LegendreImage+ with adaptive allocation.
  3. Experimental level: Main experiments and ablations demonstrate the reconstruction performance, parameter efficiency, and design trade-offs of both methods in Tables 2, 4, 6, and 7.

Specific configurations, allocation strategies, and optimization choices implement and enhance these representations and are not claimed as separate core innovations.

The statement that GFES "builds on classical localized approximation" does not fully capture the cross-method connections we identify. From 3DGS [1], 2DGH [2], and other existing methods, we inductively identify the common form of signal expansion combined with spatial localization and describe it using classical generalized Fourier expansion. Appendix F / Table 9 of the revised manuscript gives the concrete correspondences:

MethodSignal CoordinatesSignal RepresentationLocalization
GI [4] / GI++ [5] / LIG [3]Local spaceConstant RGBGaussian function
3DGS [1]Viewing directionSpherical harmonicsGaussian function
2DGH† [2]Local spaceHermite polynomialsGaussian function
WIPES (2D) [6]Modulation phaseConstant and cosineGaussian function
LegendreImage / LegendreImage+ (Ours)Local squareLegendre polynomialsCompact raised-cosine window

GI denotes GaussianImage. † The Hermite expansion in 2DGH describes the shape field before its nonlinear activation.

These correspondences span different signal domains: 3DGS expands its signal over the directional domain while using Gaussian functions for spatial localization. The cosine modulation in WIPES can be written as a finite Fourier representation on the phase domain, with Gaussian functions still providing spatial localization. The signal and localization domains can differ, yet these representations share the same structure. The locally weighted structure of LegendreImage is also mathematically related to Shepard interpolation and MLS, but this local connection does not by itself capture the unified view across signal domains.

This shared structure also provides a mathematical starting point for constructing new representations. Classical expansions such as Chebyshev, Jacobi, Zernike, and Fourier–Bessel provide signal representations associated with their respective domains and measures. Combining them with spatial localization expands the range of potential GFES instances beyond Table 9.

From this understanding, we deductively construct LegendreImage, instantiating the shared structure for local color representation on a bounded domain. The Legendre expansion describes how local color varies with position, while the compact window provides spatial localization. Fixed and adaptive allocation correspond to LegendreImage and LegendreImage+, respectively, with construction details in Sec. 3. The choice of specific basis functions is a result of this construction. Its first-order truncation spans the same scalar function space as a standard local affine model and is one configuration within this representation family, which can be extended to higher orders.

Original Table 4 shows reconstruction gains from higher-order expansion at a fixed primitive count, demonstrating per-primitive expressiveness. In revised Table 2, fixed-allocation LegendreImage improves from 42.38 dB at zeroth order to 48.29 dB at first order under the same 4M trainable-parameter budget, demonstrating the parameter efficiency of local expansion under this budget. Our use of K=1K=1 in the main experiments is motivated by parameter efficiency and does not represent a theoretical limit on the expressiveness of LegendreImage.

AuthorsOpenReview ↗

W2 & Q1: Contribution of Local Color Expansion and Execution of the Gain-Based Selector (Part I)

Part I: What are the respective roles of local color expansion, normalized aggregation, spatial localization, and adaptive allocation?

Under the same 4M trainable-parameter budget, first-order local color expansion improves parameter efficiency even with adaptive allocation disabled, and normalized aggregation also improves reconstruction quality in the base LegendreImage representation. The core of this work is primitive representation design guided by GFES, with adaptive allocation in LegendreImage+ used to further enhance performance on top of LegendreImage.

(1) Local color expansion and normalized aggregation.

The new experiments use fixed-allocation LegendreImage on the complete DIV2K validation set, with the same 4M trainable-parameter budget and raised-cosine window type. Densification and relocation are disabled. The first-order configurations use RGB-shared linear coefficients.

LegendreImage ConfigurationParams / PrimitivePrimitive CountTrainable-Parameter BudgetPSNR (dB)Data Source
K=1K=1, Normalized7571,428=⌊4,000,000/7⌋571{,}428=\lfloor4{,}000{,}000/7\rfloor4,000,00048.29Original Table 7
K=0K=0, Normalized5800,000=⌊4,000,000/5⌋800{,}000=\lfloor4{,}000{,}000/5\rfloor4,000,00042.38New in revised Table 2
K=1K=1, Additive7571,428=⌊4,000,000/7⌋571{,}428=\lfloor4{,}000{,}000/7\rfloor4,000,00044.79New in revised Table 2

The two Normalized configurations show that the first-order representation achieves higher reconstruction quality despite using fewer primitives. The two first-order configurations keep the window, primitive count, initialization rule, and optimization settings unchanged, replacing only normalized aggregation with direct summation ∑iWiQi\sum_iW_iQ_i. These experiments have been included in Sec. 5.5 and Table 2 of the revised manuscript.

Reconstruction gains from increasing per-primitive expressiveness are also demonstrated by experiments at a fixed primitive count. Appendix C / Table 4 of the original manuscript uses LegendreImage with adaptive allocation disabled, a fixed count of 100k primitives, and independent RGB coefficients. The configurations with K=0,1,2,6K=0,1,2,6 achieve 29.24, 32.53, 35.46, and 45.67 dB, respectively.

(2) Adaptive allocation.

Original Fig. 4 and Tables 6 and 7 compare LegendreImage and LegendreImage+ across the evaluated budgets on DIV2K and Kodak. The complete adaptive allocation pipeline improves reconstruction over the fixed-allocation representation. The new fixed-allocation experiments examine local color expansion and normalized aggregation independently of this adaptive enhancement.

(3) Spatial localization.

Compact support is a premise of the LegendreImage construction, while GFES itself encompasses representations with both compact and non-compact support. Given the practical constraint of evaluating primitives over a finite rendering region, we follow GFES to first specify a bounded local domain and compact localization, then construct the Legendre signal expansion on that domain to obtain LegendreImage.

The Gaussian/Hermite experiments in Sec. 5.1 / Table 2 of the original manuscript construct local signal expansions with non-compact Gaussian localization and compare them with LegendreImage+. These experiments examine the performance of different complete representation configurations under the same parameter budget within GFES. The newly added Appendix G / Table 10 further reports non-positive-definite covariance records and analyzes how local coordinates depend on covariance, providing mechanistic and empirical support for considering local signal expansion and spatial localization jointly.

Sec. 5.4 / Table 2 of the original manuscript keeps the compact support extent and other settings unchanged and compares Raised-Cosine, Cubic Smoothstep, Triangular, and Parabolic windows, obtaining 51.83, 51.77, 51.01, and 50.63 dB, respectively. This ablation examines window shape within a fixed compact support, while the cross-method experiments evaluate different complete representation configurations.

Reviewer_67FTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes GFES, where each splatting primitive represents a local function rather than a constant color. The proposed LegendreImage method uses a raised-cosine window with compact support and a first-order Legendre polynomial expansion. Each primitive has 7 trainable parameters (3 RGB, 2 shared slope, and 2 center offsets). LegendreImage+ further introduces residual-based primitive insertion and relocation. Experiments on DIV2K and Kodak show that LegendreImage+ performs worse than GaussianImage++ at lower parameter budgets but achieves higher PSNR at larger budgets.

Critical strengths

The basic idea is simple and reasonable. The paper reports training time and rendering speed, including the additional computational cost compared with GaussianImage++.

Critical weaknesses
  1. One of my main concern is the limited novelty. GFES is introduced as a general framework for representing the signal carried by each primitive using an orthogonal function expansion. However, LegendreImage uses only a first-order Legendre expansion. At this order, the Legendre basis is exactly the monomial basis {1, u, v}. The resulting local color function is simply an affine function, with no difference in representation or parameterization from an ordinary first-order polynomial.

LegendreImage combines these local affine functions using compact spatial windows and normalized weighted averaging. This is closely related to classical moving least squares and Shepard interpolation. The use of local polynomial functions with spatial weighting is not new. The main difference appears to be the application of these ideas to image splatting.

The paper emphasizes choosing an orthogonal basis according to the signal domain and measure. However, the first-order Legendre basis used in the experiments is identical to the ordinary affine basis, so the method does not demonstrate any specific benefit from Legendre orthogonality. The main contribution seems to be a combination of local affine color functions, compact support, normalized blending, and adaptive primitive allocation.

  1. The theoretical analysis provides limited support for the proposed method. Appendix A derives the approximation error for a truncated Legendre expansion of a smooth local function. This is a standard result from polynomial approximation theory and is not specific to the proposed method. The analysis only considers polynomial approximation of a smooth function. It does not cover the spatial window, blending of overlapping primitives, or adaptive primitive allocation used in LegendreImage+. It also does not address image edges and discontinuities, which can significantly affect reconstruction quality.

  2. Another main concern is that the claimed contribution is not well supported by the ablation. The paper argues that the local function space and spatial weighting should be designed jointly. However, the experiments do not clearly separate the effects of these components. In particular, GaussianImage++ with first-order Hermite terms achieves 43.86 dB at 4M parameters, worse than 48.06 dB for the original GaussianImage++. Adding first-order terms increases the function space available to each primitive, although the fixed parameter budget reduces the number of primitives. A performance drop of this big should be explained.

The comparison between Hermite-augmented GaussianImage++ and LegendreImage+ also changes several components at once, including the spatial window, blending method, primitive geometry, and allocation procedure. It is not possible to tell which of these changes is responsible for the improvement.

  1. The parameter accounting is questionable. Appendix B counts 7 trainable parameters per primitive, including 2 center offsets. However, LegendreImage+ initializes only half of its primitives on a regular grid. The remaining primitives are inserted at positions selected from reconstruction residuals. These inserted primitives have anchor positions that cannot be recovered from the original grid and must be stored. Their support sizes also depend on the insertion and relocation stages. The paper does not explain how the complete representation is stored.

The Gaussian baselines count position, covariance, and color parameters. Comparing these against LegendreImage+ without accounting for all the information needed to reconstruct its primitives is not a fair comparison of representation size. At 4M parameters, fixed-allocation LegendreImage achieves 48.29 dB, compared with 48.06 dB for GaussianImage++. The difference is only 0.23 dB. LegendreImage+ reaches 51.83 dB, but its actual storage cost is not fully reported.

  1. Several baseline results are questionable. In Table 1, GaussianImage with Cholesky parameterization drops from 40.66 dB at 4M parameters to 29.74 dB at 7.2M on DIV2K. The RS variant drops from 38.81 dB to 25.42 dB. WIPES with Cholesky parameterization also becomes worse as the parameter budget increases. Such large performance drops after increasing representation capacity should be explained and discussed in detail.

  2. The practical benefit is unclear, particularly for image compression. LegendreImage+ only outperforms GaussianImage++ when using a large number of parameters. At lower parameter budgets, GaussianImage++ performs better. This limits the usefulness of the proposed method for compact image representation.

In addition, the comparison is based on the number of trainable parameters. This is different from the actual storage size. Since the related Gaussian-based methods also target image compression, the paper should include rate-distortion comparisons using actual bitrates.

  1. Figures 1 and 5 show considerable color variation within individual Legendre primitives. However, since the first-order coefficients are shared across RGB channels, each primitive can only change its RGB values by the same amount. The paper should explain how the displayed color variations are produced, particularly since Figure 1 uses clipping to visualize the primitive fields.

The paper also reports much higher PSNR for LegendreImage+ at larger parameter budgets, while its MS-SSIM is sometimes slightly lower than GaussianImage++. This difference should be discussed. In addition, Figure 2 shows unusually large improvements in selected image regions. The authors should explain how these examples were chosen and whether they are representative of the overall results.

Questions for authors
  1. What is the actual storage cost of LegendreImage+, including inserted primitive anchors, support information, and any required metadata?

  2. Why do GaussianImage and WIPES show such large PSNR drops when the parameter budget increases?

  3. Can the authors provide controlled experiments separating the effects of local affine expansion, compact support, normalized blending, and adaptive allocation?

  4. Can the authors provide actual rate-distortion comparisons against GaussianImage++ and standard image codecs, including quantization and entropy coding?

11 comments
AuthorsOpenReview ↗

Acknowledgments

We sincerely thank the reviewer for the time and effort devoted to a thorough and detailed assessment of our work. We particularly appreciate your examination of the paper's central claims in relation to the mathematical analysis, experimental comparisons, and implementation details needed to support them. Your questions about the information required for complete reconstruction, the actual storage cost of the representation, and the consistency between quantitative metrics and visual results provided concrete directions for improving the manuscript.

These comments prompted us to reexamine the original explanations and evidence point by point, add controlled experiments and diagnostic analyses, clarify the scope of the theoretical analysis, and correct issues identified in the evaluation procedure. We greatly value this scrutiny of the method and its experimental details. It helped us explain more clearly the conditions underlying each conclusion, the results supporting it, and how readers can verify those results.

We are deeply grateful for your careful attention to these details. Following these revisions, the manuscript provides a more complete account of the method, the evidence, and its practical costs. Below, we respond to each comment and indicate the corresponding revisions.

To address the soundness concerns, we have added controlled experiments, clarified the scope of the theoretical analysis, and audited the evaluation and storage accounting. Upon acceptance, we plan to release our implementation and experimental configurations for independent verification.

AuthorsOpenReview ↗

W1: Novelty of GFES and LegendreImage

As summarized in the final paragraph of Sec. 1, our contributions span three levels:

  1. Conceptual level: We introduce GFES in Sec. 1, revealing the shared mathematical structure of different primitive representations.
  2. Methodological level: In Sec. 3, we instantiate GFES as two image representation methods, namely LegendreImage with fixed allocation and LegendreImage+ with adaptive allocation.
  3. Experimental level: Main experiments and ablations demonstrate the reconstruction performance, parameter efficiency, and design trade-offs of both methods in Tables 2, 4, 6, and 7.

Specific configurations, allocation strategies, and optimization choices implement and enhance these representations and are not claimed as separate core innovations.

Local polynomials and spatial weighting have an established mathematical foundation, and the recurrence of similar structures across primitive representations motivates our study. From 3DGS [1], 2DGH [2], and other existing methods, we inductively identify the common form of signal expansion combined with spatial localization and describe it using classical generalized Fourier expansion. Appendix F / Table 9 of the revised manuscript gives the concrete correspondences:

MethodSignal CoordinatesSignal RepresentationLocalization
GI [4] / GI++ [5] / LIG [3]Local spaceConstant RGBGaussian function
3DGS [1]Viewing directionSpherical harmonicsGaussian function
2DGH† [2]Local spaceHermite polynomialsGaussian function
WIPES (2D) [6]Modulation phaseConstant and cosineGaussian function
LegendreImage / LegendreImage+ (Ours)Local squareLegendre polynomialsCompact raised-cosine window

GI denotes GaussianImage. † The Hermite expansion in 2DGH describes the shape field before its nonlinear activation.

These correspondences span different signal domains: 3DGS expands its signal over the directional domain while using Gaussian functions for spatial localization. The cosine modulation in WIPES can be written as a finite Fourier representation on the phase domain, with Gaussian functions still providing spatial localization. The signal and localization domains can differ, yet these representations share the same structure. The locally weighted aggregation in LegendreImage is also mathematically related to MLS and Shepard interpolation.

This shared structure also provides a mathematical starting point for constructing new representations. Classical expansions such as Chebyshev, Jacobi, Zernike, and Fourier–Bessel provide signal representations associated with their respective domains and measures. Combining them with spatial localization expands the range of potential GFES instances beyond Table 9.

From this understanding, we deductively construct LegendreImage, using a Legendre expansion to represent color variation within a bounded local domain and a compact window for spatial localization, as detailed in Sec. 3.1.

The first-order truncation of LegendreImage spans the same scalar function space as a standard local affine model. The O(h2)O(h^2) projection error bound on the normalized local domain in Appendix A explains the effectiveness of affine approximation for smooth scalar signals at small spatial scales. Conversely, the affine form alone does not provide the above correspondences across signal domains or the construction path for this representation family.

In revised Table 2, fixed-allocation LegendreImage improves from 42.38 dB at zeroth order to 48.29 dB at first order under the same 4M parameter budget. Original Table 4 fixes the primitive count at 100,000 and uses independent RGB coefficients to show reconstruction gains from higher orders. These experiments respectively assess parameter efficiency and the reconstruction benefits of higher-order expansion.

AuthorsOpenReview ↗

W3 & Q3: Component Contributions and Joint Design of Color Functions and Spatial Weighting

1. Effects of local expansion, normalization, spatial localization, and adaptive allocation

Under the same 4M trainable-parameter budget, first-order local expansion improves parameter efficiency with adaptive allocation disabled, and normalized aggregation improves reconstruction quality in the base LegendreImage representation.

All configurations use fixed-allocation LegendreImage on the 100 DIV2K validation images, with Raised-Cosine windows and densification and relocation disabled:

ConfigurationParams / PrimitivePrimitive CountTrainable-Parameter BudgetPSNR (dB)Data Source
K=1K=1, Normalized7571,428=⌊4M/7⌋571{,}428=\lfloor4\mathrm{M}/7\rfloor4,000,00048.29Original Table 7
K=0K=0, Normalized5800,000=⌊4M/5⌋800{,}000=\lfloor4\mathrm{M}/5\rfloor4,000,00042.38New in revised Table 2
K=1K=1, Additive7571,428=⌊4M/7⌋571{,}428=\lfloor4\mathrm{M}/7\rfloor4,000,00044.79New in revised Table 2

The first-order Normalized configuration achieves higher quality with fewer primitives than zeroth order. The first-order variants differ only in aggregation: normalized blending or direct summation ∑iWiQi\sum_iW_iQ_i, with the same window, primitive count, initialization rule, and optimization settings. These ablations are added in revised Sec. 5.5 / Table 2.

Original Appendix C / Table 4 evaluates the fixed-allocation LegendreImage variant with 100,000 primitives and independent RGB coefficients, showing that higher truncation orders improve reconstruction without adaptive allocation.

Original Fig. 4 and Tables 6 and 7 compare LegendreImage with LegendreImage+ across the evaluated budgets on DIV2K and Kodak, showing that the complete adaptive allocation pipeline consistently improves PSNR over fixed allocation.

Compact support is a premise of the LegendreImage construction, while GFES itself encompasses representations with both compact and non-compact support. Original Sec. 5.4 / Table 2 keeps support and other settings fixed: Raised-Cosine, Cubic Smoothstep, Triangular, and Parabolic achieve 51.83, 51.77, 51.01, and 50.63 dB in LegendreImage+.

2. Hermite extensions and joint design

Under the current implementations and the 4M budget, adding first-order Hermite terms to GaussianImage++ [5] and LIG [3] does not improve parameter efficiency and is accompanied by more non-positive-definite covariance records.

Revised Appendix G / Table 10 reports new non-PD records aggregated over all 100 DIV2K validation images. Percentages use each configuration's total primitive budget as the denominator:

MethodParams / PrimitivePrimitive Budget / ImageNon-PD Records (100 Images)Records / BudgetPSNR (dB)
GaussianImage++8500,0005,663,52011.3270%48.06
GaussianImage++ with Hermite10400,0006,723,80016.8095%43.86
LIG8500,0001,192,7792.3856%44.86
LIG with Hermite10400,00011,118,73127.7968%41.07

PSNR values are from original Tables 1–2.

The reviewer notes that comparing Hermite-augmented GaussianImage++ with LegendreImage+ changes several components at once. Sec. 5.1 asks whether the same first-order local expressiveness translates into parameter efficiency in different primitive constructions. First-order Hermite and Legendre expansions are affine in their respective local coordinates. The GFES comparison therefore examines color expansion together with spatial representation.

Within-method comparisons between zero- and first-order expansions provide further evidence. With each Gaussian method's spatial parameterization, aggregation, and training pipeline unchanged, Hermite terms lower reconstruction quality. Without adaptive allocation, first-order LegendreImage improves from 42.38 to 48.29 dB despite also using fewer primitives. Thus, under a fixed budget, adding local color freedom does not automatically improve efficiency and must be assessed together with spatial representation.

Appendix G also describes an optimization coupling: Hermite local coordinates depend on covariance. For one color channel, let WiW_i be the spatial weight, QiQ_i the local color function, Φi=WiQi\Phi_i=W_iQ_i the primitive contribution, and Σi\Sigma_i the covariance. Then

∇ΣiΦi=Qi∇ΣiWi+Wi∇ΣiQi.\nabla_{\Sigma_i}\Phi_i = Q_i\nabla_{\Sigma_i}W_i + W_i\nabla_{\Sigma_i}Q_i.

For constant color, ∇ΣiQi=0\nabla_{\Sigma_i}Q_i=0. With first-order terms, this term is generally nonzero. Adding color coefficients therefore changes spatial-parameter optimization as well as expanding the local color function space. This coupling mechanism and the experimental results support joint design: greater local expressiveness must be considered together with parameter cost, spatial parameterization, and optimization.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Locations refer to the revised manuscript unless stated otherwise. W/Q labels follow our reviewer-specific responses.

1. Main Text

  1. Sec. 3.2: Clarified the validity requirements for gain-based selection and the use of residual fallback for remaining insertions. Comments: mUYj W2 & Q1.

  2. Table 1: Added fixed-allocation LegendreImage to the main comparison using results taken directly from original Tables 6–7. Comments: mUYj W2 & Q1, 67FT W3 & Q3, kBh9 (allocation novelty).

  3. Table 2 / Sec. 5.5: Added fixed-allocation ablations of local color expansion and normalized aggregation at 4M trainable parameters. The first-order normalized baseline is taken directly from original Table 7. Comments: mUYj W2 & Q1, 67FT W3 & Q3.

  4. Table 2 / Sec. 5.3: Added independent RGB at the same primitive count, explicitly marking its 6,285,708 trainable parameters as an exception to the 4M budget. Expanded the discussion of comparisons at a fixed budget and a fixed primitive count to distinguish per-primitive expressiveness from parameter efficiency. Comments: mUYj Q3 & W1.

  5. Sec. 5.1: Clarified the purpose of comparing complete GFES instances under the same budget and the rationale for joint design discussed in Appendix G. Comments: 67FT W3 & Q3, mUYj W2 & Q1.

  6. Secs. 6.3 and 7: Added a discussion of local affine models within GFES and how weighted blending can produce non-affine global reconstructions. Refocused the conclusion on the primitive design question in Sec. 1. Comments: kBh9 (novelty, Q1), mUYj W1, 67FT W1.

  7. Eq. (6): Corrected mn≠0mn\ne0 to (m,n)≠(0,0)(m,n)\ne(0,0), retaining the first-order terms. Comments: kBh9, Minor issue.

  8. Fig. 2 caption: Replaced “representative” with “selected” to describe the displayed DIV2K regions. Comments: 67FT W7.

2. Appendices

  1. Appendix A.1 / Eqs. (26)–(27): Added an analysis of error propagation for nonnegative normalized aggregation and discussed local variation near image edges. Comments: 67FT W2.

  2. Appendix D / Table 8: Added the Field computation, coordinate conventions, and coefficients needed to reproduce the illustrated primitives, identifying the Fig. 1 examples. Comments: 67FT W7.

  3. Appendix F / Table 9: Added explicit correspondences between existing primitive representations and the construction of LegendreImage from GFES. Comments: kBh9 (novelty, Q1), mUYj W1, 67FT W1.

  4. Appendix G / Table 10: Added covariance diagnostics and analysis of gradient coupling for the Hermite extensions on the full DIV2K validation set. Comments: 67FT W3 & Q3, mUYj W2 & Q1.

  5. Appendix H / Tables 11–14 / Fig. 6: Added analysis of high-budget GaussianImage degradation, building on LIG's prior findings, with error decompositions, visualizations, and a comparison with WIPES. Comments: mUYj Q2, 67FT W5 & Q2.

  6. Appendix I / Fig. 7 / Tables 15–17: Added complete storage and codec descriptions for LegendreImage and LegendreImage+, with RD comparisons against GI++ and JPEG. The Fig. 7 caption clarifies that GI++'s official rates exclude some decoding overhead. Comments: 67FT W4 & Q1, W6 & Q4.

  7. Tables 4–7: Corrected MS-SSIM evaluation errors involving mixed prediction/GT memory layouts and cuDNN TF32, including DIV8K results. Comments: 67FT W7.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.