This video explores the fundamental limitations of traditional triangle-based 3D computer graphics and evaluates new methods for representing complex scenes. After discussing Neural Radiance Fields (Nerfs) and their performance drawbacks, the speaker focuses on 3D Gaussian Splatting, highlighting its speed but also its inefficiency for flat surfaces due to "blobby" artifacts. The core of the video introduces a novel approach called "Triangle Splatting," which leverages a new type of triangle capable of both hard and fuzzy edges. Through side-by-side comparisons of rendered scenes (a Ferrari and two castles), the speaker demonstrates that Triangle Splatting consistently produces superior visual quality, particularly in preserving fine details, reflections, and distant background elements, with significantly less smearing compared to Gaussian Splatting. Despite its proven graphical superiority and high rendering speeds, the technology faces adoption challenges due to its nascent research status, lack of real-time viewers, and the existing software ecosystem's entrenched reliance on older rendering primitives.
Summarizing the Evolution of 3D Graphics Primitives
0:00 The "Triangle Problem": For 30 years, computer graphics have relied on triangles, which are inherently inefficient and fundamentally broken for representing the organic, smooth shapes of the real world, requiring millions of tiny triangles for basic realism.
1:30 Introduction of Nerfs (2020): Neural Radiance Fields offered a paradigm shift, using AI to learn light interaction and produce photorealistic outputs from multiple photos, without traditional meshes or triangles. However, Nerfs were impractically slow, taking hours to train and rendering at only 1-2 frames per second.
2:23 Rise of 3D Gaussian Splatting (2023): This method represented scenes with millions of tiny, fuzzy blobs, achieving extremely fast render rates (60-1000 FPS), making it suitable for games and mobile applications.
3:00 Gaussian Splatting's Own "Triangle Problem": Despite its speed, Gaussian Splatting struggles with flat surfaces and sharp edges, requiring millions of blobs to approximate simple planes and often resulting in smudged or "blobby" artifacts.
3:41 Introducing Triangle Splatting (2025 Research): Researchers developed a "new type of triangle" that can exhibit both hard edges and fuzzy Gaussian distributions, combining the best of both worlds. This approach leverages GPU efficiency for triangles, promising thousands of frames per second.
4:56 Practical Testing and Evaluation: The speaker downloaded and tested the "Triangle Splatting" project code, comparing its rendering quality against 3D Gaussian Splatting using standard metrics: SSIM (structural similarity), PSNR (peak signal-to-noise ratio), and LPIPs (AI perceptual similarity).
10:29 Ferrari Scene Comparison: Triangle Splatting demonstrated superior detail in reflections, sharper elements (e.g., brake caliper dots), more accurate tree reflections in paint, and significantly less smearing in distant backgrounds compared to Gaussian Splatting.
12:30 Castle Scene Comparisons: Across multiple castle scenes, Triangle Splatting consistently preserved more background detail, produced sharper overall images, and reduced smearing in clouds and foliage, making distant objects look more realistic.
17:14 Ground Truth Comparison: Triangle Splatting's renders were almost indistinguishable from the actual source images, highlighting its impressive photorealism and detail retention.
18:07 Adoption Challenges: Despite its superior quality, Triangle Splatting faces a significant hurdle: it is a new research project without a real-time viewer, and the existing software ecosystem for 3D graphics is already heavily invested in traditional triangles and is shifting towards Gaussian Splatting, potentially leaving this promising technology behind.
18:53 Call to Action: The speaker concludes by urging software developers to consider building upon Triangle Splatting, asserting that 3D Gaussian Splatting is not the ultimate solution for the future of computer graphics.