3D Gaussian Splatting achieves real-time novel-view synthesis at 1080p

Bernhard Kerbl and colleagues applied 3D Gaussian splatting to novel-view synthesis with end-to-end optimisation and adaptive density control, achieving high-quality results at 1080p resolution and real-time frame rates, without the slow neural rendering that previous approaches required.

Background

Rendering a scene from a new camera angle, given only a set of photos, is called novel-view synthesis. Radiance field methods had made this possible with impressive quality, but they relied on neural networks that were expensive to train and slow to run. Faster alternatives existed, but they traded image quality for speed.

The particular gap was at the high end: no existing method combined explicit representation, competitive training time, and real-time display rates at 1080p for complete, unbounded scenes.

What happened

Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis submitted their paper to arXiv on 8 August 2023. It was published in ACM Transactions on Graphics, volume 42(4).

The method works in three parts. Starting from sparse points produced during camera calibration, the scene is represented with 3D Gaussians. These preserve the properties that make continuous volumetric radiance fields good for optimisation, while avoiding computation in empty space. Then the system runs interleaved optimisation and adaptive density control of those Gaussians, with anisotropic covariance optimised to fit the scene accurately. Finally, a fast visibility-aware rendering algorithm handles anisotropic splatting, which speeds up both training and the final render.

The result was real-time novel-view synthesis at 1080p, at 30 frames per second or above, with visual quality the authors reported as competitive with or exceeding prior methods on several datasets by standard metrics. Competitive training times were maintained alongside the rendering speed, which previous fast methods had not managed.

The key shift was representing a scene with explicit 3D Gaussians rather than querying a neural network for every point in space. That made the rendering algorithm fast enough to run in real time without giving up the quality that radiance field approaches had established.

Why it mattered

Before this work, no method combining explicit representation, competitive training time, and 30 fps or above at 1080p on unbounded scenes had been demonstrated. The approach showed that explicit 3D Gaussian representations could match the visual quality of neural radiance fields while being fast enough to render interactively, opening a path to practical real-time use of radiance field techniques.

Sources

• 3D Gaussian Splatting for Real-Time Radiance Field Rendering. arxiv.org. Primary source
• 3D Gaussian Splatting for Real-Time Radiance Field Rendering. arxiv.org. Primary source

Cite this page

AI Achievements. (2023). 3D Gaussian Splatting achieves real-time novel-view synthesis at 1080p. Retrieved 2026-08-29, from https://achievements.ai/milestone/3d-gaussian-splatting-achieves-real-time-novel-view

@misc{achievements_3d_gaussian_splatting_achieves_real_time_novel_view,
  title  = {3D Gaussian Splatting achieves real-time novel-view synthesis at 1080p},
  author = {{AI Achievements}},
  year   = {2023},
  url    = {https://achievements.ai/milestone/3d-gaussian-splatting-achieves-real-time-novel-view}
}