Innovative Solution to a Classic Computer Graphics Problem

What is the best way to represent three-dimensional objects on-screen? This issue has been tormenting anyone in the field of computer graphics—from computer game developers to Disney animators—for over three decades. PhD student Guy Fargion of Prof. Ofir Webber’s research group offers an innovative solution to one of the key problems in this sphere: flattening 3D surfaces onto 2D planes. His solution utilizes neural networks for rapid prediction of complex optimization problem-solving.

Fargion completed his dual bachelor’s in computer engineering and physics, then continued on an accelerated track to earn his master’s in computer engineering. In his PhD project, he is researching digital geometry processing, and developed an algorithm for solving a classic computer graphics problem: surface parameterization, flattening 3D surfaces onto 2D planes. “We encounter engineering issues when transitioning from 3D to 2D and flattening or stretching geometric elements. One of the major ones is geometric distortion, which, although unpreventable, can be reduced or bounded. This distortion causes gaps between the original coloring executed in 2D and the end result in 3D,” Fargion explains. “This problem has been troubling many researchers in the field, because computers encounter it whenever they present an object on the screen, for instance a human figure in a video game, movie, or animation. When we want to present a three-dimensional object, we must colorize it or attach an image to it, but first we must flatten it and make it two-dimensional. That makes coloring far easier. Once the model has been colored, we recreate the texture over to the 3D model.”

Per Fargion, algorithms have been trying to cope with this problem as early as 30 years ago, and each year presents new breakthroughs that offer more efficient solutions. “But despite the advances, current classic methods that produce practical results and minimize distortion are often carried out in slow iterations,” he says. “Our innovation is based on neural networks. Our research shows how neural networks can be trained to predict an approximation of the classic solution; in other words, the network operates as a fast predictor for solving this complex problem of optimization. Although the neural network’s predicted solution is an approximation and slightly more distorted than the classic solution, it can be achieved relatively quickly compared to the classic methods. The reason is that the network utilizes GPUs, a tool for parallel computation, allowing the prediction for each separate triangle to take place simultaneously.”

Fargion’s algorithm, developed with the guidance of Prof. Webber, was demonstrated in the paper “Learning Metric Fields for Fast Low-Distortion Mesh Parameterizations,” which was accepted for presentation at the Eurographics 2025 conference, to take place in London this May. “The conference is being held by an organization that promotes research in the field of computer graphics, including digital geometry processing. The paper will also be published in the organization’s prestigious journal, Computer Graphics Forum,” says Fargion. I’m grateful for the opportunity to present the paper and our algorithm in this professional setting. Using neural networks for this purpose is an innovative solution that produces credible results and opens a doorway to further research in the field.”