demos

During my internship at Amazon AGI Foundations, 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.
During my internship at NEC Laboratories America, I worked on multimodal reasoning for driving video: a multimodal LLM perceives the scene, reasons about the ego vehicle's motion state, and predicts where it goes over the next few seconds.
During my internship at Amazon, I worked on Restore-R1: an image restoration agent that judges a degraded photo with a multimodal LLM and learns, through reinforcement learning, which restoration tools to call and in what order [Paper].
Degraded input: a lightning storm shot through heavy rain streaks and sensor noise.
Restore-R1 output: the same lightning photograph with the rain streaks and noise removed.
Degraded input: a glass skyscraper obscured by rain streaks and sensor noise.
Restore-R1 output: the same skyscraper with the rain streaks and noise removed.
Restore-R1 Low quality
Drag the slider to reveal the restored images.
During my internship at Adobe Research, I worked on quality-controllable visual retrieval: a lightweight language model refines a short query at a requested aesthetic and relevance level, so the quality of what CLIP-based retrieves is steered through words alone rather than by retraining the computationally expensive retrieval model [Paper] [Webpage] [Live Demo].