Creating realistic 3D shape deformations

What is it about?

3D objects are often represented as surfaces made of many small triangles. However, the same object can be divided into triangles in many different ways, which can make AI tools produce inconsistent results. We introduce a way to generate realistic 3D deformations that works reliably even when the number or arrangement of triangles changes. Our method starts with carefully designed random patterns and gradually turns them into detailed 3D forms, such as human poses and elastic rest poses on the floor. This makes it possible to create high-quality results on much larger and more detailed 3D models than those used during training.

Featured Image

Photo by Eugene Evans on Unsplash

Photo by Eugene Evans on Unsplash

Why is it important?

Most AI methods for 3D meshes work only with the same triangle layout they were trained on. Our key contribution is to make the model independent of that layout - a requirement previous approaches have largely overlooked. Our approach lets models trained on manageable meshes generate convincing deformations on models with millions of triangles, reducing the need to retrain for each version of the same shape.