What was broken
Packaging variants were produced by hand, one design pass at a time. Every copy change or regional variant meant another manual round, and brand guidelines had to be re-checked each time.
Shipped
Prompt-Driven Packaging Imagery
An image processing pipeline that takes a natural-language prompt and applies text overlay and editing to product packaging — classical computer vision and diffusion models working together rather than either alone.
RoleApplication Development Associate
Duration1 year 8 months
DomainComputer vision
StatusShipped
How it was built
The pipeline combines OpenCV for the deterministic work — placement, overlay, geometry — with Stable Diffusion for generation and the Azure Vision API for understanding what is already in the frame. Prompt templates carry the brand guidelines, so alignment is enforced by the template rather than left to whoever is writing the prompt. The image manipulation pieces are built as reusable components rather than a single script.
What came of it
Generated image quality and brand-guideline alignment both improved once prompts were templated, and the reusable components cut processing time on repeat variants.
[00][Variants produced?]
[00][Processing time cut?]
3Vision systems combined