Complex-Valued 2D Gaussian Representation for Computer-Generated Holography
Yicheng Zhan, Xiangjun Gao, Long Quan, Kaan Akşit · 2025
Read Complex-Valued 2D Gaussian Representation for Computer-Generated Holography by Yicheng Zhan, Xiangjun Gao, Long Quan, Kaan Akşit (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-15022.txt.
Introduction
Yicheng Zhan, Xiangjun Gao, and two co-authors represent holograms with soft two-dimensional Gaussian blobs that carry complex values. Computer-generated holography computes the light pattern a display must emit to float a three-dimensional image in space. Working pixel by pixel gives a huge search space, while Gaussian primitives compress the pattern into far fewer, well-shaped pieces.
The chosen primitive hits the best possible joint focus in space and frequency, a classic result from Gabor's theory, and cuts the parameter count fivefold. A custom differentiable renderer, paired with a fast GPU light propagation core, trains the representation end to end. Experiments show lower memory use, faster optimization, and clearly sharper quality scores than standard approaches.
Catalog record
| CREATOR | Yicheng Zhan, Xiangjun Gao, Long Quan, Kaan Akşit |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 60 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-15022
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Complex-Valued 2D Gaussian Representation for Computer-Generated Holography Yicheng Zhan, Xiangjun Gao, Long Quan, Kaan Akşit Abstract We propose a new hologram representation based on structured complex-valued 2D Gaussian primitives, which replaces per-pixel information storage and reduces the parameter search space by up to 10:1. To enable end-to-end training, we develop a differentiable rasterizer for our representation, integrated with a GPU-optimized light propagation kernel in free space. Our extensive experiments show that our method achieves up to 2.5 × \times lower VRAM usage and 50% faster optimization while producing higher-fidelity reconstructions than existing methods. We further introduce a conversion procedure that adapts our representation to practical hologram formats, including smooth and random phase-only holograms. Our experiments show that this procedure can effectively suppress noise artifacts observed in previous methods. By reducing the hologram parameter search space, our representation enables a more scalable hologram estimation in the next-generation computer-generated holography systems. 1 1 footnotetext: Equal contribution. † Corresponding author. 1 Introduction Holographic displays are a promising technology for realistic Three-Dimensional (3D) imaging Kim et al. 2024 , with Computer-Generated Holography (CGH) providing the computational means for generating high-fidelity holograms for these displays. Unlike natural images, which record only light intensity, holograms capture intensity, interference, and diffraction phenomena. As shown in Figure 1 , holograms exhibit markedly different spatial characteristics. Therefore, a key challenge in CGH is to design compact and efficient representations that preserve high-frequency details of holograms while supporting scalable optimization Wang et al. 2022 ; Shi et al. . Figure 1 : Comparison between a natural image and different hologram formats. Unlike the smoother pixel structures in natural images, holograms produce dense high-frequency and random spatial variations that are challenging to represent. (Source Image: pom ) Conventional image representation methods, such as Implicit Neural Representation (INR) Sitzmann et al. 2020 ; Rumelhart et al. 1986 , optimize a continuous implicit function to represent an image. However, as implicit functions favor continuous low-frequency data, they are unable to faithfully capture high-frequency details of the hologram. In parallel, autoencoder-based approaches Bohan 2023 ; Wang et al. 2024 ; Preechakul et al. 2022 ; Song et al. 2023 which are typically pretrained on natural images, also tend to be suboptimal as their learned priors do not generalize well to hologram structures. Additionally, emerging Gaussian-based image representations Zhang et al. 2024 ; Zhang et al. 2025 ; Zeng et al. 2025 have recently been proposed for natural image modeling. However, these methods directly encode hologram pixels and neglect modeling interference and diffraction phenomena, making them unsuitable for the ideal solution for compact hologram representation. In light of this limitation, we propose a new hologram representation based on complex-valued 2D Gaussians. Naively assigning per-pixel parameters to random-valued holograms will produce unstructured high-frequency patterns that resist compact parameterization. In contrast, our representation models the hologram with structured Gaussian primitives and explicitly incorporates light propagation. To enable end-to-end optimization, we develop a differentiable rasterizer for rendering complex-valued 2D Gaussians, integrated with a GPU-optimized light propagation kernel. In practice, our method reduces the parameter search space by up to a 10:1 ratio, decreases VRAM usage by up to 2.5 \times , and accelerates optimization by 50% , while outperforming existing methods in reconstruction fidelity. Despite the challenges of finding compact hologram representation, existing random-valued hologram methods Kavaklı et al. 2023b ; Kavaklı et al. 2023a ; Kuo et al. 2023 ; Schiffers et al. 2025 ; Chao et al. 2024 which typically rely on dense, per-pixel optimization, still suffer from image quality issues. These methods assign independent degrees of freedom to each pixel, which greatly expands the search space and hinders convergence toward better solutions. Such suboptimal optimization often manifests as noise artifacts in reconstructed holograms. To mitigate these artifacts, encoding schemes Hsueh and Sawchuk 1978 and learning-based CGH methods Shi et al. 2022 ; Peng et al. 2020 ; Shi et al. 2021 ; Zhan et al. 2024 have been explored to smooth hologram to enhance image quality Peng et al. 2025 . However, as shown in Figure 2 , this smoothing restricts pupil freedom, color contrast, and brightness Schiffers et al. 2023 ; Choi et al. 2021 ; Kuo et al. 2023 , which limits the quality of practical displays. Motivated by these limitations and by the need for practical display deployment, we propose a conversion procedure that adapts our representation to different hologram formats, including Smooth Phase-only Hologram (POH) and Random POH. Our extensive experiments demonstrate that this procedure can significantly suppress the noise artifacts observed in prior methods Kavaklı et al. 2023b ; Kavaklı et al. 2023a . Our contributions are summarized as follows: • We propose a new hologram representation based on complex-valued 2D Gaussians , reducing the parameter search space by up to 10:1 while preserving fidelity. • We develop a differentiable rasterizer with GPU-optimized light propagation, enabling efficient and scalable hologram optimization and rendering. • We design a conversion procedure that adapts our representation to practical hologram formats and significantly reduces noise artifacts observed in previous methods. For readers less familiar with the underlying principles of CGH, we present the essential concepts in Sec. 2.2 . 2 Related Work 2.1 Gaussian and Learned Image Formation Recent advances in natural image representation have explored compact alternatives to dense pixel-wise parameterizations. Autoencoder-based methods Preechakul et al. 2022 ; Bohan 2023 ; Wang et al. 2024 and INRs Sitzmann et al. 2020 ; Saragadam et al. 2023 ; Lindell et al. 2022 ; Rumelhart et al. 1986 map natural images into latent spaces or continuous functional representations, enabling structured compression but often favoring smooth, low-frequency content. More recently, building on 3D Gaussian Splatting (3DGS) and its variants Kerbl et al. 2023 ; Gao et al. 2025 ; Mallick et al. 2024 ; Wu et al. 2024 , several works have extended Gaussian primitives from neural rendering to natural image encoding Zhang et al. 2024 ; Zeng et al. 2025 ; Zhang et al. 2025 , leveraging 2D Gaussians for efficient image rendering and compression. While these Gaussian-based methods show promising result on representing natural images, they neglect modeling interference and diffraction phenomena, making them unsuitable for directly representing holograms. In this paper, we aim to bridge this gap by proposing a complex-valued 2D Gaussian-based hologram representation that explicitly integrates light propagation during optimization, achieving both parameter search space reduction and high-fidelity representation for computer-generated holography. 2.2 Preliminary Concepts: CGH CGH is a computational imaging task that performs wave-based rendering, synthesizing holograms to reconstruct 3D scenes for holographic displays. Unlike natural images that capture only smooth, low-frequency intensity variations, a hologram simultaneously encodes light’s intensity, interference, and diffraction. As shown in Figure 1 , this wave-optical data format manifests as dense, high-frequency, and random-valued structures that challenge conventional image representations. Holographic data can be represented as a complex hologram: \mathbf{H}=A\cdot\exp(j\varphi) with amplitude A and phase \varphi , requiring specialized modulators; or as a phase-only hologram: \mathbf{H}_{\text{POH}}=\exp(j\varphi) , compatible with commercial holographic displays. POH variants include Smooth POH with spatial multiplexed phase Hsueh and Sawchuk 1978 and Random POH with directly optimized phase Schiffers et al. 2023 . See Sec. 3.4 for more details of different hologram formats. Figure 2 : (Simulated) Smooth hologram shows high