Using Composable Diffusion and AI-generated Images to Elevate Personalized Customer Experiences
INTRODUCTION
AI has hit the news and its a great time for us to share some thoughts on composable diffusion. With artificial intelligence feeling increasingly less artificial, ideas that previously seemed far fetched last now seems perfectly plausible. At its most basic, composable diffusion involves splitting images into smaller parts and analyzing them separately. It’s a term commonly used in machine learning, to describe the combination of multiple text descriptions for image generation. And a foundation for text to image generation.
MIRROR MIRROR ON THE WALL
So why is this of interest? We think this has the potential to be a composable commerce, digital asset game changer. In an age where customer experience is key, what if an intelligent fitting room could understand what an occupant wanted by responding to user-initiated cues? Delivering images of available or customized products on the fly, tailored to that unique user experience. “Mirror, I like these jeans but show me this red crew with a v-neck and add a pocket.” High street retail seems a long way off from delivering high-volume customized clothing, but this is not impossible. It might even reduce waste. Advancements in this field and the increasing demand for personalised shopping experiences could lead to a new age of innovation. Enabling retailers to gain a more comprehensive understanding of customer preferences and tailor their offerings accordingly.
MIT's Computer Science and Artificial Intelligence Laboratory has been at the forefront of research with a number of groundbreaking studies published in 2020. Researchers developed a deep learning model that could be applied to the automatic identification of clothing styles from customer images, with the ability to categorize them. An AI model trained by one large dataset of images to distinguish one style from another. Very much future thinking of course but information that could potentially be used to create a highly personalized shopping experience for customers, by recommending clothing items that are similar to their preferred style. In another MIT study, researchers focused on using composable diffusion to improve the accuracy of object recognition in images. A trained deep learning model capable of identifying specific items such as shoes, bags, and accessories. The model was able to accurately identify the specific item in an image, even when it was partially obscured or placed in an unusual position.
A shiba inu wearing a beret and black turtleneck
Hierarchical Text-Conditional Image Generation with CLIP Latents - OpenAI 13/04/2022
Accelerating rapidly, investment comes as the demand for AI-powered content generation grows, alongside increasing demand for advanced and flexible methods of generating images from natural descriptions. In September 2022 the London-based startup behind AI image generator Stable Diffusion raised $101 million in a funding round to support custom versions of Stable Diffusion for users at a larger scale.
AI & IMAGE OWNERSHIP
Whilst AI image generation has the potential to revolutionize many industries, including retail this new technology is not without its challenges, one of the biggest is the issue of copyright. In August 2022 a Colorado fine arts competition awarded a prize to AI-generated artwork, drawing further attention to the issue. When an AI model generates from a dataset, a unique digital asset is created and the question of who owns the rights to the image is raised. Participating factors include the role of the model generating the image, the role of the person or company providing the natural language inputs, and the role of the person or company using the image. A complex and controversial issue, not one for this think tank article but certainly one to consider for dataset owners and users alike. Of course, if you already own the rights to a large image dataset then many of these ownership questions are voided. As an aside, the US Copyright Office has issued guidelines stating that works created by machines or algorithms are not considered original works unless they are generated by a human author. So who owns that red v-neck crew? And how does this law sit for businesses operating across borders?
AI & BIG BROTHER
Let's explore the potential for AI models to learn new information without forgetting previously learned knowledge. This is a crucial area of research in the field of AI, as it has the potential to significantly enhance the capabilities of AI systems. The idea is certainly possible, not yet the reality. In theory a model could incorporate mechanisms for retaining previously learned knowledge, alongside ones that allow it to learn new information. A continual learning framework where a model is trained on a sequence of tasks, each of which builds upon the knowledge acquired in previous tasks. Retail could potentially benefit from this research in several ways. For example, an AI model with the ability to retain and build upon previous knowledge could be used to generate images for product advertisements. The model could be trained on a wide range of product categories, and as new products are introduced, learn, retain and generate images that are more accurate and relevant to a user.
AI & HUMAN WORKING MEMORY
Working memory is a cognitive system that allows us to temporarily store and manipulate information for problem solving and decision making. It’s highly sophisticated, dynamic and influenced by many factors, including attention, motivation, and prior experience. For AI models, previously learned knowledge. AI models can perform tasks that are similar to human working memory, but it’s currently not possible for them to fully simulate the complexity, flexibility and diversity required.
For retailers, working memory plays an important role in customer experience. For example, recalling and processing information about products, remembering past purchases and retaining information about promotions and deals. When working memory capacity is limited, difficulties arise with complex decision making and processing multiple pieces of information at once. Leading to frustration or dissatisfaction with the shopping experience. Retailers can consider these factors when designing store layouts, creating marketing campaigns, and providing customer support mechanisms that enhance the shopping experience.
While AI models can perform tasks that require the use of a temporary memory store, such as image captioning or machine translation they cannot replicate it. A limited amount of information is stored and typically there is no ability to allocate resources dynamically based on task demands. Or respond to changing conditions in real time. AI models have made significant progress in performing tasks that are similar to human working memory, but they still have a long way to go. A question remains, should this go anywhere?
CONCLUSION
In conclusion, we have more questions than when we started writing! Composable diffusion certainly does have the potential to revolutionize the way retailers gather information and change the way in which personalized shopping experiences are created. We haven’t touched on search but potential applications exist for what users view as the humble website search box. Behind the box lay a complex web of queries and unseen technologies, delivering the information requested. As the business of search becomes increasingly more intelligent, AI has the potential to make it much more intuitive too. Granular to the preferred button, collar, cuff and hem length. Research is just the beginning. We expect to see many more advancements in the future.