Karan Desai

I am a founding member at World Labs. I work on curating petabyte-scale data as well as model pre-training for Atlas, our camera-conditioned generative model. In the past, I have also contributed to earlier generation models powering Marble. World Labs is joining AMD, I will be starting my new role soon.

I finished my PhD in Computer Science at the University of Michigan in May 2024, advised by Justin Johnson. My PhD work centered on visual representation learning, vision-language models, and image segmentation. I am glad to have enjoyed working on these topics in academia back then, before they became mainstream in the current wave of Gen AI products since the virality of ChatGPT release in 2022. My thesis, titled Language Supervision for Computer Vision, is available publicly.

These days, my favorite pockets of time at work are working with data. I enjoy the process of iteratively growing large multimodal datasets, increasing their quality density by hand-designing data transforms to filter or select samples to better train generative models. I like finding simple ways to source data, be it from the internet or by manually recording videos. I like vibe coding bespoke web interfaces to manually hand-annotate samples or to simply eyeball thousands of images. Sometimes, I sit down and eyeball tens of thousands of images and video clips to verify quality and correctness; in my opinion, improving data quality is the best bang for the buck to improve generative models. This process is exhausting, yet very satisfying.