Complex-Valued 2D Gaussian Representation for Computer-Generated Holography

TL;DR

This work introduces a compact, differentiable hologram representation built from complex-valued 2D Gaussian primitives that explicitly model light propagation. By parameterizing holograms with a small set of structured Gaussians and integrating a differentiable rasterizer with a GPU-accelerated propagation kernel, the approach achieves up to a 10:1 reduction in parameter search space, significant VRAM savings, and faster optimization while attaining higher fidelity reconstructions. The authors also propose a conversion pipeline to practical phase-only formats (Smooth and Random) that suppresses noise artifacts and preserves structural details. Across simulations and hardware experiments, the method outperforms prior Gaussian-based and learned CGH methods in both fidelity and efficiency, providing a scalable pathway toward next-generation holographic systems. The work further demonstrates real-time potential via depth- and distance-variant visualizations and outlines concrete directions for real-time video, multiplexing, and camera-in-the-loop extensions.

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 2.5x 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.

Figures (19)

• 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: pomegranate)
• Figure 2: (Simulated) Smooth hologram shows high quality at the pupil center but degrade severely with pupil shifts, whereas random hologram remains visible. (Source Image: Wilson2009)
• Figure 3: Hologram reconstruction via free-space light propagation. (Source Image: Burger2014)
• Figure 4: Overview of our pipeline. Complex-valued 2D Gaussians are rasterized into a complex hologram (amplitude and phase), which is propagated to multiple depth planes using optimized light propagation. Reconstructions are compared with RGB+D derived targets at different focal distances, and we report PSNR, SSIM, and LPIPS. Here, Light Prop denotes light propagation. (Source Image: Burger2014)
• Figure 5: Runtime (bar) and VRAM usage (line) across spatial resolutions for our method ($L = 3$), comparing CUDA-based with the PyTorch baseline. Red downward arrows and percentages indicate the VRAM reduction rate.