Bazaart, a popular photo editing app, with over 5 million users, recently upgraded its AI Magic Backgrounds feature. The work required collaboration between Deep Learning data scientists, prompt engineers, and product managers, and continuous iterations on training and testing models, prompts, and other configurations. In the past, rolling out production-grade features could take up to 2 months, a timeframe that a leading app cannot afford in the Gen AI space.
Magic Backgrounds (replacing the background of an object) is one of the most popular features introduced by the Gen AI revolution. It is used by both businesses and consumers, for e-commerce, entertainment, photo editing, and other purposes.
Bazaart has the best data scientists, prompt engineers, artists, and the know-how to train a superior model. However, they lacked a streamlined process for comparing different model outputs, running multiple test cases, and gathering feedback from the team.
Due to this, the decision to roll out a new model wasn’t made based on data but on anecdotal evidence.
To address this problem we had to give Bazaart two superpowers:
Creating hundreds of images with different parameters should be a matter of minutes not hours, and accessible to anyone on the team.
Analyzing and comparing thousands of images needs to be an easy task if you want to launch fast.
First, we connected Bazaart’s image generation pipeline to our platform and created a powerful form that allows anyone in the organization to generate images in bulk.
Generated images are automatically saved in the platform along with the parameters used to create them. We can then provide advanced tools for viewing and rating images, analysis, and insights.
Working with images as the core work entity is challenging with normal communication solutions such as Slack, emails, and Google Drive, which is why all our platform provides one place to view, discuss, and store photos.
The effect of generating hundreds of images with a click of a button and analyzing them in minutes allowed Bazaart to quickly iterate and gain confidence in what works best.
In the future, they will use these generated photos as benchmarks for comparing new models and different prompting.
Working with Magic Grids enabled Bazaart to test out and decide to release their new model in less than a week. Before that releasing these kinds of features could take up to 2 months.
For companies that generate images in production, Magicflow provides a framework for working with images, a methodology for testing and releasing new features, improves output quality, and saves valuable time.
Give it a try: https://console.magicflow.ai
Or book a meeting with the founders: https://calendly.com/magicflowai/meet-team
