Stanislaw Szymanowicz: Generating Interactive 3D Scenes in Seconds. скачать в хорошем качестве

Stanislaw Szymanowicz: Generating Interactive 3D Scenes in Seconds.

The Talk given by Stanislaw Szymanowicz KUIS AI Talks on December 16,2025 𝐓𝐢𝐭𝐥𝐞: Generating Interactive 3D Scenes in Seconds 𝐀𝐛𝐬𝐭𝐫𝐚𝐜𝐭: Our community has made great progress in generative modelling, enabling image and video generation at quality we have not seen before. However, the majority of generated media does not permit interaction, a crucial capability for exciting application areas such as gaming or embodied AI. In this talk, I will argue that feed-forward generation of explicit 3D scenes is the most viable direction towards generating interactive media at scale. I will present Bolt3D, a method capable of generating entire 3D scenes from any number of photos in just 6 seconds on a single GPU. The user can then interact with the output scenes in browser in real-time, paving the path towards real-time interaction with generated worlds. I will conclude the talk with a discussion of future directions for unified 3D reconstruction and 3D generation systems such as Bolt3D. Short Bio: Stan Szymanowicz is a final-year PhD student at VGG, University of Oxford and a Visiting Researcher at Meta in London. He interned at Google San Francisco, Microsoft Research Cambridge and received his Masters with Distinction from the University of Cambridge. He researches the intersection of Generative Models and 3D Computer Vision.