Artist directed non-photorealistic rendering framework

Non-photorealistic rendering (NPR) is an emerging approach in computer graphics, where visual content is synthesized not trying to replicate physical interactions between light and materials, but also enabling the expression of an artistic style or intent. Historically, this has been a field dominated by image processing algorithms.

Currently the strongest trend being AI powered techniques, which have made their way into mobile devices and real-time camera filters.

This style of approach focuses on transforming any image into an art piece, with no need for user input, simply providing a guidance original image, or general descriptions of the visual primitives (e.g. brush strokes, line drawings, color palette). However, precise user input is essential for expressing artistic intent.

When pursuing a strong art direction or vision with non-photorealistic rendering, we must give artists control over the rendering process If non-photorealistic rendering has the goal of expressing artistic intent.

Previous work that focused on providing artist control of certain shading aspects using an object-based approach has demonstrated success in creating results that can be used in toon-shaded animated content (e.g. games and animated movies) and creating results that are both temporally consistent and still capture the artists vision with limited input.

This is a challenging scenario for image-space techniques.

The cost of training neural network models that considers both time consistency and allows for user parameters to be provided is currently not feasible.

A procedural solution that clearly exposes parameters of the rendering process would better fit production pipelines of entertainment studios and existing tools.

This project will focus on developing novel NPR techniques that support artistic expression through flexible controls of an NPR pipeline, allowing for a variety of styles, algorithms, and media using an object-based approach and a custom rendering pipeline.