Up-res Deformation Field Learning for Highly Detailed Data-driven Liquids

The objective of the proposed project is to procedurally increase the visible resolution of a coarse liquid simulation within a machine learning pipeline for the visual effects (VFX) industry. We introduce an optical flow-based learning approach to model the deformations between a coarse input liquid simulation and its equivalent at a higher resolution. The deformation fields used as a training set are computed from multiple examples of low- and high-resolution signed distance fields (SDF). An interpolation step allows us to synthesize a new and up-res version of the coarse input liquid from the matching and learned deformation model. Our interpolation scheme also provides a way of improving user-defined and localized regions of the input simulation. This approach offers an efficient and reusable tool to enhance the apparent fine details of a liquid simulation for VFX.

Faculty Supervisor:

Pierre Poulin

Student:

Partner:

Technical University of Munich

Discipline:

Computer science

Sector:

Technology; Entertainment and Media

University:

Université de Montréal

Program: