Disney Enterprises, Inc.

Eth Zurich (eidgenossische Technische Hochschule Zurich)

Marco Manzi

Marios Papas

Christopher Richard Schroers

Tunc Ozan Aydin

Thijs Vogels

David M Adler

Farnood Salehi

Henrik D Dahlberg

Gerhard Röthlin

Andre C Mazzone

Per H Christensen

Melvin Ott

Zhilin Cai

Mark A Meyer

A system includes a hardware processor and a system memory storing software code and one or more machine learning (ML) models. The hardware processor is configured to execute the software code to train a first ML model of the one or more ML models as a denoising feature selector, generate, using the trained first ML model a plurality of candidate feature sets, and identify a best volumetric feature set of the plurality of candidate feature sets using a predetermined selection criterion. The hardware processor is further configured to execute the software code to train, using the identified best volumetric feature set, one of the first ML model or a second ML model of the one or more ML models as a denoiser, receive an image including noise due to rendering, and denoise, using the trained denoiser, the noise due to rendering to produce a denoised image.

Volume denoising with feature selection

Certain aspects of the present disclosure provide techniques for adaptive sampling for rendering using deep learning. This includes receiving, at a sampler in a rendering pipeline, a plurality of rendered pixel data, wherein the sampler includes a first machine learning (ML) model. It further includes generating a sampling map for the rendering pipeline using the first ML model and the plurality of rendered pixel data, including predicting a plurality of pixel values in the sampling map based on a generated distribution of pixel values. It further includes rendering an image using the sampler, the sampling map, and a denoiser in the rendering pipeline.

Adaptive sampling using deep learning

Techniques are disclosed for training and applying a denoising model. The denoising model includes multiple specialized denoisers and a generalizer, each of which is a machine learning model. The specialized denoisers are trained to denoise images associated with specific ranges of noise parameters. The generalizer is trained to generate per-pixel denoising kernels for denoising images associated with arbitrary noise parameters using outputs of the specialized denoisers. Subsequent to training, a noisy image, such as a live-action image or a rendered image, can be denoised by inputting the noisy image into the specialized denoisers to obtain intermediate denoised images that are then input, along with the noisy image, into the generalizer to obtain per-pixel denoising kernels, which can be normalized and applied to denoise the noisy image.

Techniques for robust image denoising

Growing community of inventors

Zurich, Switzerland

Marco Manzi