Computer Vision Model Development for a Generative AI Application

Client

Client - A generative AI photo editing app for marketers and sellers

Industry

Mobile application, photo editing, marketing

Services

State of the art model development

Description

A photo editing app that was used by more than 1 million users, developed a image beautification and generation use case. They had in house expertise and had developed a good model in multiple iterations over the past 2 years.

They reached out to us to improve this model further and be better than their competitors.

Solution

1. We first did a analysis of the state of the art models and architectures

2. We setup a data enrichment and annotation pipeline - Created a scalable data definition, improved annotation guidelines

3. We setup a regular cadence to test the best models with updated data pipeline

For this project we heavily relied on frameworks like pytorch, paddle-paddle and pytorch lighnting for multi-gpu training.

For infrastructure we relied on 4xA100s , 4xT4 and 2x3090

In terms of scale of data we were operating at 20-50 thousand finely annotated image samples for each

Result

After a deep dive of model architectures, we were able to develop a custom architecture that was

1. Better than state of art research publications in 2022-23

2. Better than 90% of their competitors in majority of use cases