Personalized Fashion Ads
In a group of four, we were able to create an multi-model pipeline that reproduces product photos to align and be styled for user data Starting from one product image, the concept reimagines the product through the lens of each audience. We will generate ads that feel personally relevant and relatable.
By adjusting visual styling elements based on each consumer’s preferences - we aim to generate personalised ads that can widen customer bases
Technical Overview
Our pipeline starts with an input seed—a product image and a descriptive text prompt. Using Grounding DINO for text-based object detection, we identify the target element within the image. The Segment Anything Model (SAM) then generates precise segmentation masks of that element.
Next comes the personalisation seed. We define user groups and craft a prompt template tailored to each group’s preferences. This personalization is applied through Stable Diffusion inpainting, which restyles the detected product region according to the audience’s tastes.
Finally, we produce the output—a set of personalized, high-quality product images saved and ready for use in targeted advertisements. This process transforms one product image into multiple audience-specific visuals that feel relevant and relatable.
Final Presentation
*
Final Presentation *
View our final presentation on how one product image can be endlessly personalized for every audience.