For Toy Story 4 I worked in the Global Technology department where I worked on several different software engineering tasks. Most notably, I worked on a core team that helped to deploy the deep learning denoiser to the studio, and I co-developed a model aggregation and visualization system for set dressing.

A team based out of Disney Research Zürich lead the development and publication of Kernel-Predicting Convolutional Networks for Denoising Monte Carlo Renderings, which was then tested at Pixar for production before being deployed as the primary denoiser for Toy Story 4 and shows to come. I joined the group at Pixar responsible for this testing and deployment, where I performed several tasks.

I wrote a back end C++ module that used the Tensorflow C-API to read the network weights, divide the input image into sub-regions, run the network on each sub-region, and then stitch the image back together. I also worked on building and deploying Tensorflow to the Pixar machines, which is much easier said than done. This required building several Tensorflow versions that would be dynamically switched depending on the client machine’s instruction set and hardware capabilities.

Unfortunately there aren’t any film frame comparisons out there, but check out Thijs Vogel’s website where he demonstrates some of the results on Cars 3 and Coco images.

One of the main sets in this show is the antiques mall, which is filled from top to bottom with unique props and models of various types and sizes. Set dressing this large environment proved to be a difficult task, so fellow GT technical director, Zach Repasky, and I built out a "model warehouse" system for organizing and viewing models. Built inside of USD, this system would scan the entire model catalog, sort models in several different ways, spatially organize them inside of USD view, and provide links that would load the model in the set dresser's active Maya session. This made it much easier for set dressers to find models of a certain size, color, or type, and easily add them to the set.