VibePhysics: Realistic Walkthrough with Blender Animation
A demo of turning real-world capture into a navigable 3D scene — training-free 4d recon+indexing system in smooth blender navigation system. [code]
I am founder and CEO of MIMI AI in London, working on GenAI systems for physical data, multimodal agents, and video world simulation.
Previously I was a senior research scientist at Meta Reality Labs, a senior LLM research engineer at Robin AI, and an applied scientist at Amazon. Before that I worked at Scape Technologies and Academia Sinica.
I received my Ph.D. in computer science from National Taiwan University. My research focuses on spatial intelligence, synthetic data generation, 3D scene understanding, image retrieval and matching, pose estimation, visual descriptors, and applied multimodal systems.
A demo of turning real-world capture into a navigable 3D scene — training-free 4d recon+indexing system in smooth blender navigation system. [code]
A lightweight framework for turning real-world videos and images into 3D maps, Blender scenes, and physical simulations — bridging feedforward reconstruction, sparse mapping, robotics, and physics in one CPU-friendly workflow.
MIMI AI, 2025 [code]
100k indoor data from scratch. Including layout, material, trajectories, along with 1000+ GPU-hours rendering results. Enabling pointcloud based VLM pre-training and post-training.
Project Aria, 2024 [dataset]
Tsun-Yi led cross domain teams including legal, product, and engineering along with collaboration with Anthropic and AWS team. Post-training outperforming OpenAI o1-preview in legal domain with extensive analysis. Efforts including RLHF, synthetic data, fine-tuning.
CVPR Workshops 2022 [paper]
CVPR 2022 [paper]
ECCV Workshops 2022 [paper]
arXiv 2020 [paper]
CVPR 2019 [paper]
ICCV Workshops 2019 [paper]
IJCAI 2018 [paper]
ICCV 2017 [paper]
CVPR 2016 [paper]