Evolutive Rendering Models

The landscape of computer graphics has undergone significant transformationswith the recent advances of differentiable rendering models. These renderingmodels often rely on heuristic designs that may not fully align with the finalrendering objectives. We address this gap by pioneering evolutiverendering models, a methodology where rendering models possess the ability toevolve and adapt dynamically throughout the rendering process. In particular,we present a comprehensive learning framework that enables the optimization ofthree principal rendering elements, including the gauge transformations, theray sampling mechanisms, and the primitive organization. Central to thisframework is the development of differentiable versions of these renderingelements, allowing for effective gradient backpropagation from the finalrendering objectives. A detailed analysis of gradient characteristics isperformed to facilitate a stable and goal-oriented elements evolution. Ourextensive experiments demonstrate the large potential of evolutive renderingmodels for enhancing the rendering performance across various domains,including static and dynamic scene representations, generative modeling, andtexture mapping.