Criteo Internship
Published:
Context: internship at Criteo AI Lab, where I worked on the AdBooster project. It was co-hosted by the CAML Core Embeddings and AIR Generative models teams at Criteo, Paris, FR
In the context of online advertising, the idea is to improve the visual quality of a product image, often originally with a plain white background, by generating an aesthetically pleasing background, conditionally to a text prompt. This process is called outpainting as opposed to inpainting, where the goal is to fill in missing parts of an image.
During the internship, I built a dataset, trained several large diffusion models for image background generation, and implemented a pipeline to generate high-quality images.
The models are largely based on the ControlNet architecture, and it’s extension:
