For an online retailer, coming up with new matchups that will have the best impact on an end customer is a tiresome and time consuming process. Looklet’s Style Creator software combines advanced technology with fully automated high-quality photography sessions for online fashion retailers. Since speed and efficiency are at the core of Looklet's offering, they are keen on always finding ways to accelerate this process even further and continue delivering high quality at great speed.
Using AI to help customers create matching outfits
Company
Looklet
AI method
Image similarity
Industry
Retail
02/ The challenge
03/ The opportunity
With the help of deep learning, the online retailers can get recommendations of how to style the models and also get suggestions for similar garments and looks to use. This not only makes the styling more efficient, but also opens the opportunity to present similar garments in case there is a shortage of an article in their stock. It also enables retailers to search their existing inventory using just an image.
04/ The background
E-commerce businesses need to ensure that their online shop is constantly relevant, showcasing the newest styles to attract customers. However, putting together an outfit from numerous articles as well as planning and setting up photoshoot after photoshoot both take a lot of time, and are very costly.
Looklet's solution to this is to help fashion retailers and brands create high quality on-model imagery. Their product offers a virtual dressing room for e-commerce companies that enables them to put together different looks on models and style them, even after the photoshoot has already taken place. This makes it easy for retailers and brands to show off new products at speed and at scale.
The way it works is that Looklet sets up an industrial studio at a location of choice to capture images of the clothes that retailers and brands sell. Then, the retailers and brands are able to create different looks in Looklet’s software, where they can choose from many or several different models in various poses and facial expressions. All the clothes from the photoshoot as well as images from previous photoshoots are stored and available to use for new style creations by the retailers.
05/ How the AI model works
By using hundreds of thousands of images and looks created in the past, a deep learning model helps find similar items to a particular target image and creates a list of styling combinations of those items. The foundation is an image similarity model. The model converts each garment image into a vector representation that is then used to calculate similarity (cosine similarity). When a customer selects a target garment and a criterion they’re trying to match for example shape, colour, and so forth, the target garment is also converted into a vector representation. After comparing the vector representation of the target image with all other images, the closest matches are returned.
Fashion and styling are highly subjective areas, but AI applied to the massive high-quality data we have opens totally new possibilities for our product, and in the end, the stylists at our clients’ retail stores can discover unique new combinations with the help of AI.
Robert Ahlborg
CPO & Co-founder, Looklet
06/ Data requirements
For this model to be effective, you need good images of clothes. Since Looklet’s images are of a very high quality, resolution was never a problem. The images were instead resized to 1024x1024 pixels, which is sufficient for the training of the chosen deep learning model. Data annotation is not needed for this model, since finding similarity is done by using the content of the image rather than metadata attached to the pictures.
07/ Results and next steps
Preliminary results have shown how image similarity can be effective at finding garments that are similar in color or shape without the need for metadata attached to each item, such as category, color, name, description, designer, and so on. The images below show two initial results of similarity-based solely on colour (top) and similarity-based solely on shape (bottom).
The next steps of this project are to fine-tune the similarity models and then add the styling combinations.
Image similarity offers many opportunities like the ones described in this customer story. Retailers and e-commerce companies in other branches can have a direct positive impact when image search is added to their workflows and websites such as increasing the speed of workflows, creating more efficient processes, and providing better accuracy in search results. This in turn could lead to companies being able to achieve results at greater speed and also enable them to scale their operations much faster.