demos
Interactive Video World Model — Amazon AGI Foundations.
During my Applied Scientist internship with the Amazon AGI Foundations team (Summer 2026), I worked on an interactive video world model: a video diffusion model that generates future frames conditioned on the incoming action, so the world can be steered step by step rather than sampled in one shot.
An action-controlled rollout from the video world model.
- Data and pipeline. Built a large-scale Unreal Engine video dataset and the accompanying training pipeline for action-controlled video generation.
- Scaling. Scaled training from a 1.3B to a 14B bidirectional video diffusion model on 32 NVIDIA B200 GPUs.
- Causal rollout. Transformed the bidirectional model into a causal architecture so frames can be generated autoregressively, which is what makes the model interactive.
- Fast inference. Applied few-step diffusion distillation to accelerate sampling.