Adobe, Inc.

Jun-Yan Zhu

Richard Zhang

Jingwan Lu

Elya Shechtman

Shabnam Ghadar

Ratheesh Kalarot

Niloy Jyoti Mitra

Zhixin Shu

Cameron Smith

Wei-An Lin

Baldo Antonio Faieta

Nathan Aaron Carr

Oliver Wang

Aliakbar Darabi

Yijun Li

Krishna Kumar Singh

Yannick Hold-Geoffrey

Gaurav Parmar

William Peebles

Alyosha Efros

The present disclosure relates to systems, non-transitory computer-readable media, and methods for utilizing machine learning models to generate modified digital images. In particular, in some embodiments, the disclosed systems generate image editing directions between textual identifiers of two visual features utilizing a language prediction machine learning model and a text encoder. In some embodiments, the disclosed systems generated an inversion of a digital image utilizing a regularized inversion model to guide forward diffusion of the digital image. In some embodiments, the disclosed systems utilize cross-attention guidance to preserve structural details of a source digital image when generating a modified digital image with a diffusion neural network.

Utilizing regularized forward diffusion for improved inversion of digital images

Systems and methods train an encoder neural network for fast and accurate projection into the latent space of a Generative Adversarial Network (GAN). The encoder is trained by providing an input training image to the encoder and producing, by the encoder, a latent space representation of the input training image. The latent space representation is provided as input to the GAN to generate a generated training image. A latent code is sampled from a latent space associated with the GAN and the sampled latent code is provided as input to the GAN. The GAN generates a synthetic training image based on the sampled latent code. The sampled latent code is provided as input to the encoder to produce a synthetic training code. The encoder is updated by minimizing a loss between the generated training image and the input training image, and the synthetic training code and the sampled latent code.

Supervised learning techniques for encoder training

An improved system architecture uses a Generative Adversarial Network (GAN) including a specialized generator neural network to generate multiple resolution output images. The system produces a latent space representation of an input image. The system generates a first output image at a first resolution by providing the latent space representation of the input image as input to a generator neural network comprising an input layer, an output layer, and a plurality of intermediate layers and taking the first output image from an intermediate layer, of the plurality of intermediate layers of the generator neural network. The system generates a second output image at a second resolution different from the first resolution by providing the latent space representation of the input image as input to the generator neural network and taking the second output image from the output layer of the generator neural network.

Multi-scale output techniques for generative adversarial networks

Utilizing cross-attention guidance to preserve content in diffusion-based image modifications

Systems and methods train and apply a specialized encoder neural network for fast and accurate projection into the latent space of a Generative Adversarial Network (GAN). The specialized encoder neural network includes an input layer, a feature extraction layer, and a bottleneck layer positioned after the feature extraction layer. The projection process includes providing an input image to the encoder and producing, by the encoder, a latent space representation of the input image. Producing the latent space representation includes extracting a feature vector from the feature extraction layer, providing the feature vector lo the bottleneck layer as input, and producing the latent space representation as output. The latent space representation produced by the encoder is provided as input to the GAN, which generates an output image based upon the latent space representation.

Direct regression encoder architecture and training

Systems and methods dynamically adjust an available range for editing an attribute in an image. An image editing system computes a metric for an attribute in an input image as a function of a latent space representation of the input image and a filtering vector for editing the input image. The image editing system compares the metric to a threshold. If the metric exceeds the threshold, then the image editing system selects a first range for editing the attribute in the input image. If the metric does not exceed the threshold, a second range is selected. The image editing system causes display of a user interface for editing the input image comprising an interface element for editing the attribute within the selected range.

Attribute control techniques for image editing

An improved system architecture uses a pipeline including a Generative Adversarial Network (GAN) including a generator neural network and a discriminator neural network to generate an image. An input image in a first domain and information about a target domain are obtained. The domains correspond to image styles. An initial latent space representation of the input image is produced by encoding the input image. An initial output image is generated by processing the initial latent space representation with the generator neural network. Using the discriminator neural network, a score is computed indicating whether the initial output image is in the target domain. A loss is computed based on the computed score. The loss is minimized to compute an updated latent space representation. The updated latent space representation is processed with the generator neural network to generate an output image in the target domain.

Techniques for domain to domain projection using a generative model

Systems and methods generate a filtering function for editing an image with reduced attribute correlation. An image editing system groups training data into bins according to a distribution of a target attribute. For each bin, the system samples a subset of the training data based on a pre-determined target distribution of a set of additional attributes in the training data. The system identifies a direction in the sampled training data corresponding to the distribution of the target attribute to generate a filtering vector for modifying the target attribute in an input image, obtains a latent space representation of an input image, applies the filtering vector to the latent space representation of the input image to generate a filtered latent space representation of the input image, and provides the filtered latent space representation as input to a neural network to generate an output image with a modification to the target attribute.

Attribute decorrelation techniques for image editing

Embodiments are disclosed for generative image congealing which provides an unsupervised learning technique that learns transformations of real data to improve the image quality of GANs trained using that image data. In particular, in one or more embodiments, the disclosed systems and methods comprise generating, by a spatial transformer network, an aligned real image for a real image from an unaligned real dataset, providing, by the spatial transformer network, the aligned real image to an adversarial discrimination network to determine if the aligned real image resembles aligned synthetic images generated by a generator network, and training, by a training manager, the spatial transformer network to learn updated transformations based on the determination of the adversarial discrimination network.

Growing community of inventors

San Jose, CA, United States of America

Jun-Yan Zhu