by Adam Pease
This week Adobe pushed further into the computer vision market with its new research on directional generative adversarial networks (DGANS). While the name may sound confusing, these AI systems are closely related to the neural networks that power the most commonly used computer vision applications, such as facial recognition. In this blog, we unpack Adobe’s new research and discuss what it means for the future of marketing and content creation.
What Are GANs?
A GAN is a Generative Adversarial Network, which is a form of a neural network for image generation. A close cousin of the convolutional neural networks used in image recognition applications, GANs are similar to other neural networks, but for an important difference. Instead of understanding and interpreting a scene, a generative AI is tasked with understanding a scene and then generating a similar scene. Rather than simply recognize visual content, GANs create visual content.
A GAN works by intaking a large training data set and then using large amounts of compute power to ‘learn’ from this data about different image categories and the relationships between images. From this, a GAN-based system can be given keywords or other instructions that direct it to try and produce an image or video that matches the content. For instance, a GAN trained on a large data set of 20th century historical paintings may be able to produce a convincing replica when it is fed the keyword ‘Picasso.’
Adobe’s foray into GANs for content automation signals the growth of the computer vision market.
Generative Content for the Enterprise
GANs grew famous years ago after Google’s DeepDream algorithm wowed the Internet with its strange, dream-like images. Since then, researchers have come a long way to bring GANs from a tool for creating bizarre pictures into an enabling technology with real enterprise solution potential. But while the applications of computer vision and scene recognition are more obvious for businesses, what is the value of generating visual content?
Realistic Human Figure Generation for Marketing
In this environment, the value of being able to automatically generate visual content begins to become more obvious. Adobe’s recent paper used a proprietary neural network architecture called a directional GAN (DGAN), which achieved high fidelity results in generating believable human figures wearing a variety of different clothes. Its approach mirrors the path followed by several startups in the computer vision space, such as rosebud.ai and DataGrid, which automatically generate human subjects.
Will Adobe jump into GAN?
Adobe has been a pioneer when it comes to providing tools for content creation for digital marketing. Adobe’s suite of applications and platforms is used to create images and videos for marketing teams around the world. Now, as marketing shifts into an era that is driven by search engine optimization and social media management, the volume of visual content required to run a campaign has increased. Aragon suspects that Adobe may dive into GANs via acquisition and rosebud.ai, DataGrid, and others could be targets.
The future of GANs and Image Generation
In the near future, clothing labels may opt to generate their clothing models through a computer vision application platform rather than take the time to pay a physical model and photographer for a shoot. Marketing teams may come to rely on application platforms that enable the partial or complete generation of their content, perhaps using a combination of templates and GAN-based procedural generation frameworks. In general, content creators may find that certain aspects of the workflow—such as the creation of primitive assets or models—are capable of being automated through AI.
Bottom Line
The future of content is generative, but it will still take some time before we get there. Adobe’s new research sheds light on what one of the tech titans of the digital marketing world is looking to bring to the table in the near future. GANs have come a long way in a short time, and there is no doubt that if an affordable and scalable solution for generating content without human input is made available to the enterprise it will bring about a sea change in the world of marketing and content creation more generally.