Russian Researchers Find Faster Way to Build Compact 3D Models and Scenes
Scientists report a method that speeds up rendering of complex 3D models while dramatically shrinking their data footprint.
At a research seminar hosted by the Artificial Intelligence Center at Moscow State University, scientists presented new techniques that significantly accelerate the rendering of complex 3D models and reduce storage requirements. The team focused on alternative geometry representations built with neural network methods.
Math Instead of Polygons
Instead of relying on traditional triangle meshes, the researchers use an SDF representation – a method that describes an object through a distance function to its surface rather than a collection of polygons. This approach allows the system to determine ray-surface intersections more quickly and reduces the number of calculations required during rendering.
More Complex the Model – Less Space It Takes
The developers also employ hierarchical structures, including BVH and octrees. These divide a scene into multiple levels and allow the renderer to access only the portions involved in a given frame. As a result, memory usage and GPU load drop substantially.
Detailed models – such as sculptures, corals, or engineering objects – can be compressed from hundreds of gigabytes down to just a few megabytes. Image precision remains high, while frame rates increase severalfold. The rendering pipeline makes more efficient use of modern GPUs and completes ray tracing faster.
Not Just for Games
The team also focused on ray tracing of NURBS surfaces, which are widely used in industrial design. Researchers combined hierarchical scene partitioning, approximate SDF models, and stochastic selection of the initial point for Newton’s method.
In test scenes of varying complexity – from a simple wine barrel to intricate industrial models – performance improved by a factor of three to six without sacrificing quality. The results have formed the basis of a series of papers now being prepared for publication and presentation at international conferences on computer graphics and computer vision, including SIGGRAPH and ECCV.
