Writing about my research in computer vision, multi-agent systems, and AI for social good to make the research more accessible.
One autoregressive diffusion model that generates coordinated human motion — for any task, any group size, any duration. We introduce MAGNet, a unified framework tackling the full spectrum of multi-person motion generation.
Does knowing your partner's movements help predict your own future motion? We tackle this question using 30 hours of swing dancing data and show that social conditioning cuts prediction error by 52%.
A deep dive into how racial composition of datasets impacts generative image models — from truncation tricks that erase minority representation to annotator biases that distort quality metrics.