Daniel Watson

@watson_nn

Research Scientist @GoogleDeepMind. Opinions my own. 🇵🇦

Montreal, Quebec danielwatson6.github.io Joined October 2017

614 Following 1K Followers 96 Posts

Latest posts 5

I've been wanting to get to writing more often, so I started a little blog attached to my personal website.

My first blog post is a tutorial on building TPU flash attention from the ground up. Let me know what you think!

danielwatson6.github.io/fa/fa.html

danielwatson6.github.io TPU Flash Attention From Scratch A tutorial on how and why flash attention works, building a flash attention kernel step by step for TPUs with JAX and Pallas, and tuning it to match splash attention.

Daniel Watson @watson_nn Jan 28

A much needed step in world modeling. By generating 360° panos, world models will truly track everything happening around them, serving as a better form of memory.

Huge shoutout to @Dazitu_616 for spearheding this awesome work. It was amazing to work with you!

Ziyi Wu @Dazitu_616 Jan 28

📢Introducing 360Anything, our method for lifting any perspective image or video to gravity-aligned 360° panoramas without using any camera or 3D information. This enables consistent novel view synthesis and 3D scene reconstruction.

Project page: 360anything.github.io

🧵

View quoted post on X

Daniel Watson @watson_nn Jul 11, 2024

[[THREAD]] Happy to announce 4DiM, our diffusion model for novel view synthesis of scenes! 4DiM allows camera+time control with as few as one input image.

Joint work with @srbhsxn* @lala_yi_li* @taiyasaki @fleet_dj
*equal contribution

Daniel Watson @watson_nn Apr 3, 2024

Very excited to share VIDIM, our diffusion model for video interpolation. Cheap and high quality video diffusion will become a reality. I have also believed for a while now that using pixels as inputs is severely underrated-- there is so much more signal in there v.s. text.

Janne Kontkanen @jannekontkanen Apr 3, 2024

VIDIM: Video Frame Interpolation Using Diffusion Models. Watch out for our generative frame interpolation magic in CVPR 2024. With Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Holynski, and Ben Poole arxiv.org/abs/2404.01203

View quoted post on X

Daniel Watson @watson_nn Oct 2, 2023

Why use RLHF when u can backprop through the sampler with any differentiable reward🥹🥹🥹

Paul Vicol @PaulVicol Oct 2, 2023

Check out @clark_kev’s and my paper on fine-tuning diffusion models on differentiable rewards! We present DRaFT, which computes gradients through diffusion sampling. DRaFT is efficient & works across many reward functions.

With @kswersk, @fleet_dj

arXiv: arxiv.org/abs/2309.17400

View quoted post on X