CV / ML researcher working on designing and building world models - both from the generation and estimation side. Currently at Google DeepMind.
Cambridge, MA
Joined August 2020
- Tweets4
- Following39
- Followers94
- Likes18
Have you ever wondered how to get a "world model" that generates a scene in the REAL ACTUAL WORLD?
See CityRAG!
Beyond just being a cool demo (you can actually be in NYC!), it hints at the future of generative video games -- where hybrid representations stabilize the generator.
Introducing CityRAG!
We wanted video generative models to be grounded in the real world — if I’m in London, I want to look around and actually see Big Ben.
CityRAG generates videos of cities featuring real buildings and roads, with arbitrary weather, people, and cars. 1/N
page: cityrag.github.io
paper: arxiv.org/abs/2604.19741
I think this idea is really powerful. What is the best way to leverage both label data as well as self-supervised techniques?
Define a good, flexible internal representation (4DGS) that allows both to help! See UFO-4D!
Some really cool work!!!
Can dense 2D supervision help learning 4D?
Introducing UFO-4D, feedforward dense 4D reconstruction from two unposed images. By leveraging 2D signals to bridge sparse 4D datasets, we enable high-fidelity 4D interpolation in one forward pass.
🔗 ufo-4d.github.io
#ICLR2026
Charles Herrmann retweeted
Can dense 2D supervision help learning 4D?
Introducing UFO-4D, feedforward dense 4D reconstruction from two unposed images. By leveraging 2D signals to bridge sparse 4D datasets, we enable high-fidelity 4D interpolation in one forward pass.
🔗 ufo-4d.github.io
#ICLR2026
We're very excited to present a new hybrid memory version of feed-forward geometric reconstruction!
The core intuition is that our architectures should be designed with type of training data we have available in mind.
The result is very long (kilometer-scale) reconstruction!!
𝗢𝗻𝗲 𝗺𝗲𝗺𝗼𝗿𝘆 𝗰𝗮𝗻’𝘁 𝗿𝘂𝗹𝗲 𝘁𝗵𝗲𝗺 𝗮𝗹𝗹.
We present 𝗟𝗼𝗚𝗲𝗥, a new 𝗵𝘆𝗯𝗿𝗶𝗱 𝗺𝗲𝗺𝗼𝗿𝘆 architecture for long-context geometric reconstruction.
LoGeR enables stable reconstruction over up to 𝟭𝟬𝗸 𝗳𝗿𝗮𝗺𝗲𝘀 / 𝗸𝗶𝗹𝗼𝗺𝗲𝘁𝗲𝗿 𝘀𝗰𝗮𝗹𝗲, with 𝗹𝗶𝗻𝗲𝗮𝗿-𝘁𝗶𝗺𝗲 𝘀𝗰𝗮𝗹𝗶𝗻𝗴 in sequence length, 𝗳𝘂𝗹𝗹𝘆 𝗳𝗲𝗲𝗱𝗳𝗼𝗿𝘄𝗮𝗿𝗱 inference, and 𝗻𝗼 𝗽𝗼𝘀𝘁-𝗼𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻.
Yet it matches or surpasses strong optimization-based pipelines. (1/5)
@GoogleDeepMind @Berkeley_AI