Canva Pty Ltd

Stefan Paul Sietzen

Sanchit Sanchit

Alexander Tack

David Fankhauser

Methods of training a machine learning model for image processing are described, as well as methods for image processing to determine a predicted rotation value for an image based on the trained machine learning model. A method of training includes utilising as a learning objective a reduction or minimisation of a classification loss. Training image pairs may be associated with classifications corresponding to rotation values for the purpose of determining the classification loss. An image may be divided into a plurality of images for determining an input to the machine learning model, for training or for image processing.

Image rotation

Described embodiments generally relate to a computer-implemented method for performing inpainting. The method includes accessing a first image; receiving a selected area of the first image; identifying a foreground area of the first image; generating a merged mask based on the union of the user selected area and the foreground area; performing an inpainting process on the area of the first image corresponding to the merged mask to generate a second image, being an inpainted image; generating a reduced mask based on the user selected area reduced by the foreground area; and combining the first image with the area of the second image corresponding to the reduced mask to produce an output image.

Systems and methods for performing inpainting

Some embodiments relate to a method of performing prompt-based image editing. The method includes accessing an image; receiving a selected area of the image; receiving a prompt, wherein the prompt is indicative of an editing instruction; generating a latent by transforming the image into visual noise; predicting, based on the latent and the prompt, a noise image corresponding to the latent; subtracting at least a portion of the noise image from the latent to generate an updated latent; generating a noisy representation of the image; generating a masked latent based on the noisy representation, the updated latent and selected area, wherein the masked latent comprises the updated latent in the areas corresponding to the selected area, and the noisy representation in the areas that do not correspond to the selected area.

Growing community of inventors

Salzburg, Austria

Stefan Paul Sietzen