CoherentRaster: Efficient 3D Gaussian Splatting for Light Field Displays Conference: Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers; July 19–23, 2026; Los Angeles, CA, USA Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers (SIGGRAPH Conference Papers ’26), July 19–23, 2026, Los Angeles, CA, USA DOI: 10.1145/3799902.3811217 ISBN: 979-8-4007-2554-8/2026/07 CCS: Computing methodologies Rendering
Gyujin Sim, Seungjoo Shin, Hosung Jeon, Gwangsoon Lee, Hyon-Gon Choo, Sunghyun Cho · 2026
Read CoherentRaster: Efficient 3D Gaussian Splatting for Light Field Displays Conference: Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers; July 19–23, 2026; Los Angeles, CA, USA Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers (SIGGRAPH Conference Papers ’26), July 19–23, 2026, Los Angeles, CA, USA DOI: 10.1145/3799902.3811217 ISBN: 979-8-4007-2554-8/2026/07 CCS: Computing methodologies Rendering by Gyujin Sim, Seungjoo Shin, Hosung Jeon, Gwangsoon Lee, Hyon-Gon Choo, Sunghyun Cho (2026) 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-2605-04509.txt.
Introduction
Gyujin Sim and five co-authors speed up rendering for light field displays, screens that show many viewpoints at once so a scene looks three-dimensional without glasses. Producing the interlaced multi-view image is expensive, and a popular fast rendering method for single views becomes costly when extended to so many angles.
Their framework renders at the level of individual subpixels. It reuses attributes across neighbouring viewpoints instead of recomputing them. It also remaps the work so graphics hardware handles the interlaced layout efficiently. Together these steps aim for real-time, high-quality light field synthesis on consumer-grade hardware. Earlier acceleration methods either wasted computation or used heavy intermediate representations.
Catalog record
| CREATOR | Gyujin Sim, Seungjoo Shin, Hosung Jeon, Gwangsoon Lee, Hyon-Gon Choo, Sunghyun Cho |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.04509
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
CoherentRaster: Efficient 3D Gaussian Splatting for Light Field Displays Conference: Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers; July 19–23, 2026; Los Angeles, CA, USA Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers (SIGGRAPH Conference Papers ’26), July 19–23, 2026, Los Angeles, CA, USA DOI: 10.1145/3799902.3811217 ISBN: 979-8-4007-2554-8/2026/07 CCS: Computing methodologies Rendering Gyujin Sim, Seungjoo Shin, Hosung Jeon, Gwangsoon Lee, Hyon-Gon Choo, Sunghyun Cho Abstract . Light field displays (LFDs) require rendering an interlaced image that encodes many view - dependent observations. This multi - view requirement introduces substantial computational overhead, making real - time rendering difficult to achieve. While 3D Gaussian Splatting (3DGS) is efficient for single-view rendering on 2D displays, directly extending it to LFDs is computationally expensive. Moreover, prior accelerations either suffer from GPU inefficiency under spatially incoherent subpixel layouts or rely on computationally heavy multi-plane intermediates. In this paper, we propose CoherentRaster , a 3DGS-based light field rendering framework that performs subpixel-level rasterization. Our method employs Cross-view Coherent Attribute Reuse to eliminate redundant computation across neighboring viewpoints and applies View-coherent Remapping to restore warp-level memory efficiency degraded by the interlaced subpixel layout. Together, CoherentRaster provides an efficient pipeline for real-time, high-quality light field synthesis on consumer-grade hardware. The code is at https://github.com/sgj0402/coherent-raster. Figure 1. Through-the-lens comparison of rendered light field image. We visualize the rendered content on a physical Light Field Display, captured from left and right viewpoints. Comparing Full-frame rendering 3DGS and CoherentRaster, our method achieves significantly higher frame rates (FPS) while maintaining visual quality. ( Garden and Bicycle scenes from the Mip-NeRF 360 dataset) teaser † † cc-license: by 1. Introduction Light field displays (LFDs) provide glasses-free autostereoscopic 3D visualization with continuous motion parallax, enabling immersive viewing experiences for 3D content without wearable devices. Unlike conventional single-view 2D displays, LFDs render interlaced light field images to simultaneously deliver view-dependent visual outcomes across a finite range of viewing angles. As LFDs with high spatial and angular resolution have recently become commercially available, there exists a growing demand for high-quality 3D content tailored to these 3D displays ( 16 ; 14 ; 22 ) . Recent advances in radiance field representation ( 17 ; 9 ) have opened up their potential to serve as effective 3D content for LFDs. Specifically, their ability to represent complex 3D scenes and support high-quality view synthesis significantly lowers the barrier to 3D content creation for everyday users. Among them, 3D Gaussian Splatting (3DGS) ( 9 ) has emerged as a practical solution for real-time, high-quality novel viewpoint rendering, achieving hundreds of frame rates on conventional 2D displays through efficient tile-based rasterization. However, applying 3DGS directly to light field displays is challenging, as LFDs require simultaneously synthesizing many view-dependent images rather than a single viewpoint. This fundamental difference introduces new computational bottlenecks that conventional 3DGS pipelines are not designed to handle. To enable real - time, high - resolution 3DGS rendering on LFDs, an LFD - specialized rasterization pipeline is required. Traditional light field reconstruction pipelines render full images for all target viewpoints and then interlace them into a single light field image ( 19 ; 2 ) . However, synthesizing these dense viewpoints is extremely costly: the computation and memory usage grow linearly with the number of required views, and the high resolution of interlaced images further amplifies the burden ( 21 ) . In practice, LFDs often require tens to hundreds of viewpoints at 2K or higher resolution, making naïve multi - view rendering prohibitively slow and preventing interactive performance. These challenges underscore the need for rendering strategies tailored specifically to 3D displays. To accelerate light field rendering, DirectL ( 24 ) first introduced subpixel-level rendering, leveraging the observation that an interlaced light field image uses only a small subset of subpixels from different viewpoints, as shown in Fig. 2 . Instead of contributing all pixels from all views, the final image is formed by interlacing only these selected subpixels. Consequently, rendering full images for every viewpoint performs substantial computation on pixels that never appear in the output. DirectL avoids this wasted work by sampling rays only for the subpixels that actually contribute to the interlaced image, bypassing full multi - view rendering. Building on this idea, 8 extend subpixel - level rendering to 3DGS by evaluating only the Gaussians required for the visible subpixels. However, applying this strategy to 3DGS introduces a new challenge: because adjacent subpixels often originate from different viewpoints, the GPU loses the spatial coherence it typically relies on. Modern GPUs execute threads in lockstep groups called warps (32 threads), making coherent memory access crucial for high throughput. When neighboring threads access disjoint view-dependent Gaussian attributes, warp - level efficiency degrades significantly. Another promising direction to efficient light field rendering would be to leverage cross - view coherence , which refers to the fact that neighboring viewpoints often observe similar scene content. Specifically, by exploiting the concept of multi - plane images, Kim et al. ( 11 ) introduce depth - aligned plane representations that allow neighboring viewpoints to share intermediate rendering results. However, this formulation faces a critical scalability bottleneck for high - resolution LFDs. Achieving high - fidelity rendering often requires hundreds of planes to suppress depth discretization artifacts, an overhead that grows rapidly with spatial resolution. As we further analyze in the experimental section, this dense intermediate representation leads to substantial computational costs, making real - time rendering impractical on consumer - grade hardware. In this paper, we introduce CoherentRaster , a real - time 3DGS - based light field rendering framework that directly addresses the key limitations of prior approaches. First, we adopt subpixel - level rasterization to evaluate only the subpixels that contribute to the interlaced light field image, eliminating the unnecessary work inherent in full multi - view rendering. Second, we exploit cross - view coherence by reusing Gaussian attributes that vary smoothly across views, substantially reducing per - view computation. Finally, we resolve the GPU inefficiency caused by the spatially incoherent subpixel layout through a view - coherent remapping strategy that reorganizes threads to improve memory coalescing and warp - level execution. Together, these components form a lightweight and scalable pipeline that leverages the benefits of cross - view coherence, similar in spirit to MPI - based methods, while avoiding the heavy intermediate representations required by multi - plane approaches, enabling real - time light field synthesis on commodity GPUs. We evaluate CoherentRaster on both synthetic and real-world benchmarks adapted for light field rendering. CoherentRaster achieves up to 23 FPS for 4K ( 3840\times 2160 ) interlaced outputs for real-world 3D scenes with 71 viewpoints on an RTX 5090 GPU. This demonstrates real-time rendering capabilities while maintaining comparable visual quality, as shown in Fig. 1 . Our contributions are summarized as follows: • We introduce CoherentRaster, an efficient 3DGS-based light field rendering framework that enables real-time, high-resolution 3D visualization on LFDs. • We propose Cross-view Coherent Attribute Reuse that shares cross-view Gaussian evaluations to reduce redundant computation across adjacent viewpoints. • We present View-coherent Remapping that reorders thread-to-subpixel mapping to restore warp-level memory efficiency under interlaced subpixel layouts. 2. Related Work 2.1. Novel View Synthesis Novel view synthesis aims to