Now let’s talk about Triangle Splatting — a new approach to real-time rendering for radiation fields.
The story is cyclical: after experimenting with semi-transparent ellipses and soft primitives, it’s time to return to the classics — triangles. And here they are, stepping onto the scene: Triangle Splatting for real-time rendering.
And this isn’t some attempt by students from China, but major players: DeepMind, Oxford, and other established teams. All of this at incredible speed — over 2400 frames per second.
This technique enables the creation of very high-quality visual effects with the ability to quickly synthesize new viewpoints, similar to Gaussian Splatting. But the main difference is that scenes are represented by triangles. While Gaussian primitives are soft and often blur details and lose fine features (for example, under a bench or near doors), Triangle Splatting preserves sharp edges and vividly renders the smallest nuances.
For those interested in the technical details: a revolution has occurred in graphics thanks to models like Neural Radiance Fields or 3D Gaussian Splatting, which have displaced triangles from their position as the primary representation in photogrammetry. We believe that triangles deserve a comeback. Therefore, we developed a differentiable renderer that directly optimizes their parameters using end-to-end gradients. The key is that each triangle is rendered as a differentiable splat. This combines the simplicity and efficiency of traditional triangles with the flexibility of adaptive density and independent primitives. As a result, compared to popular Gaussian splatting methods, our approach shows better accuracy, faster convergence, and significantly increased rendering speed.
Triangles are, after all, a classic: they are easy to integrate with standard graphics stacks, work well with GPUs, and are very efficient. For example, for the “Garden” scene, we achieved a speed of over 2400 FPS at a resolution of 1280×720 on an RTX 4090, using an existing mesh renderer. These results demonstrate that triangle-based representation is well-suited for fast and high-quality scene synthesis. It is similar to traditional computer graphics but enhanced with modern differentiable frameworks, allowing for mesh-level optimization.
It seems that the guys have given a new life to Pixar’s micro-polygons: during rendering, we accumulate the contribution of all projected triangles, using alpha blending from front to back. And since all these steps are differentiable, we can fine-tune the triangle parameters through gradient-based learning.
It will be interesting to see how this works together with Unreal Engine and Nanite.
Their documentation explicitly states: Triangle Splatting combines differentiable scene optimization with classical graphics pipelines. The entire triangle structure is compatible with any mesh-based renderers, ensuring seamless integration into existing systems. In game engines, it achieves over 2400 FPS at 1280×720 resolution — all on an RTX 4090 graphics card.
Photogrammetry seems to have come full circle and is returning to its roots.
Code? Easy: https://github.com/trianglesplatting/triangle-splatting