20230162407. HIGH RESOLUTION CONDITIONAL FACE GENERATION simplified abstract (Adobe Inc.)

HIGH RESOLUTION CONDITIONAL FACE GENERATION

Organization Name

Adobe Inc.

Inventor(s)

Ratheesh Kalarot of San Jose CA (US)

Timothy M. Converse of San Francisco CA (US)

Shabnam Ghadar of Menlo Park CA (US)

John Thomas Nack of San Jose CA (US)

Jingwan Lu of Santa Clara CA (US)

Elya Shechtman of Seattle WA (US)

Baldo Faieta of San Francisco CA (US)

Akhilesh Kumar of San Jose CA (US)

HIGH RESOLUTION CONDITIONAL FACE GENERATION - A simplified explanation of the abstract

This abstract first appeared for US patent application 20230162407 titled 'HIGH RESOLUTION CONDITIONAL FACE GENERATION

Simplified Explanation

The present disclosure is about systems and methods for image processing. It describes an image processing apparatus that can generate modified images by changing attributes or landmarks of an input image. The apparatus uses a machine learning model to encode the input image and obtain a joint conditional vector that represents the attributes and landmarks in a vector space. This joint conditional vector is then modified to create a latent vector, which is used to generate the modified image. The machine learning model is trained using a generative adversarial network (GAN) with a normalization technique, followed by joint training of a landmark embedding and attribute embedding to reduce inference time.

• The disclosure describes an image processing apparatus that can modify images by changing attributes or landmarks.
• The apparatus uses a machine learning model to encode the input image and obtain a joint conditional vector.
• The joint conditional vector is modified to create a latent vector, which is used to generate the modified image.
• The machine learning model is trained using a generative adversarial network (GAN) with a normalization technique.
• Joint training of a landmark embedding and attribute embedding is performed to reduce inference time.

Potential Applications

This technology has potential applications in various fields, including:

• Face editing and modification in digital photography and image editing software.
• Virtual reality and augmented reality applications, where synthetic faces can be generated and modified in real-time.
• Facial recognition systems, where attributes and landmarks of a face can be modified to enhance or improve recognition accuracy.

Problems Solved

This technology addresses several problems in image processing:

• It provides a method to generate modified images by changing attributes or landmarks, allowing for easy editing and modification of images.
• The use of a machine learning model and generative adversarial network improves the quality and realism of the modified images.
• The joint training of landmark and attribute embeddings reduces inference time, making the image processing faster and more efficient.

Benefits

The use of this technology offers several benefits:

• It allows for easy and efficient modification of images by changing attributes or landmarks.
• The generated modified images are of high quality and realism.
• The reduced inference time improves the speed and efficiency of image processing.

Original Abstract Submitted

the present disclosure describes systems and methods for image processing. embodiments of the present disclosure include an image processing apparatus configured to generate modified images (e.g., synthetic faces) by conditionally changing attributes or landmarks of an input image. a machine learning model of the image processing apparatus encodes the input image to obtain a joint conditional vector that represents attributes and landmarks of the input image in a vector space. the joint conditional vector is then modified, according to the techniques described herein, to form a latent vector used to generate a modified image. in some cases, the machine learning model is trained using a generative adversarial network (gan) with a normalization technique, followed by joint training of a landmark embedding and attribute embedding (e.g., to reduce inference time).