My wife (left) and I (right) at the Mendocino arch.

Vighnesh Birodkar

I am a machine learning researcher in the space of video generative models. I am deeply interested in compression, real-time, and long context generation. I have been generating videos since 2017. I believe that profound insights come from running a lot of simple experiments. I currently work in the video pre-training team at Luma AI. For my previous work experience, see my Resume.

Selected Research

Sample what you can't compress

We train auto encoders with diffusion loss (without GANs) and show that we can achieve both better compression and higher quality generation.

arXiv

Gemini 2.5 and Veo

Google's flagship AI model. I added video data and metrics to the Gemini codebase and trained a prototype diffusion transformer that, when scaled up, became Veo 1.

Gemini technical report Veo

Videopoet: A large language model for zero-shot video generation

We train an LLM with discrete video tokens and show that it is a powerful framework for video generation.

arXiv Website

The surprising impact of mask-head architecture on novel class segmentation

By fixing a critical flaw in Region Proposal networks, we show that we can get strong mask generationzation across classes by using a simple and deep mask head.

arXiv Website

Semantic redundancies in image-classification datasets: The 10% you don't need

We show that you can throw away 10% of Imagenet without any loss in evaluation accuracy.

arXiv

Open source

Anchor free object detectors in TF Object Detection API.
scikit-learn project-template and Elkan's k-means algorithm.
Region Adjacency Graphs and blob-detection in scikit-image.
FreDo Editor, a frequency domain image editor.

Undergraduate projects

Visual Programming Editor for Arduino boards.
AC Circuit Simulator for the Indian government's ek-Shiksha programme.
A maze-solving robot for NIT Surathkal Automaton 2012.
An omnidirectional robot for NIT Surathkal Automaton 2011.