Donny Y. Chen

After a year of team work, we're thrilled to introduce Depth Anything 3 (DA3)! 🚀 Aiming for human-like spatial perception, DA3 extends monocular depth estimation to any-view scenarios, including single images, multi-view images, and video. In pursuit of minimal modeling, DA3 reveals two key insights: 💎 A plain transformer (e.g., vanilla DINO) is enough. No specialized architecture. ✨ A single depth-ray representation is enough. No complex 3D tasks. Three series of models have been released: the main DA3 series, a monocular metric estimation series, and a monocular depth estimation series. The core team members, aside from me: @HaotongLin, Sili Chen, Jun Hao Liew, @donydchen. 👇(1/n) #DepthAnything3

Depth Anything 3 Recovering the Visual Space from Any Views

How can an AI model learn the underlying dynamics of a visual scene? We're introducing Trajectory Fields, a new way to represent video in 4D! It models the path of each pixel as a continuous 3D trajectory, which is parameterized by a B-spline function of time. This unlocks powerful physical AI tasks: ✨ Video Motion estimation and understanding. 🔮 Scene Motion forecasting. 🎯 Goal-conditioned trajectory generation. 🚀 Dynamic fusion: recovering the complete object while observing it is moving. Find more details in our work "Trace Anything: Representing Any Video in 4D via Trajectory Fields" This work was led by our amazing intern @xinhangliu123 (xinhangliu.com).

Excited to share our latest work from the ByteDance Seed Depth Anything team — Trace Anything: Representing Any Video in 4D via Trajectory Fields 💻 Project Page: trace-anything.github.io 📄 Paper: huggingface.co/papers/2510.13… 📦 Code: github.com/ByteDance-Seed… 🤖 Model: huggingface.co/depth-anything… 🖐️ Interactive Results: trace-anything.github.io/viser-client/i…

Interesting ICLR submissions 🤩 Depth Anything 3 - My TLDR: Init multi view transformer of VGGT with later layer DINO weights and use teacher model trained on synthetic data only for pseudo labelling real world datasets openreview.net/forum?id=yirun… Trace Anything - My TLDR: VGGT like model predicting N view geometry and motion as a trajectory field represented using splines and control points openreview.net/forum?id=BqaCh… The field is evolving very fast!

🎉ZPressor has been accepted to @NeurIPSConf 2025. 🎊The code and model checkpoints have also been open sourced on Github and Huugging Face. Welcome to star and use it! Code: github.com/ziplab/ZPressor Project Pages: lhmd.top/zpressor Models: huggingface.co/lhmd/ZPressor

VolSplat Rethinking Feed-Forward 3D Gaussian Splatting with Voxel-Aligned Prediction

Had a fantastic experience working with Chuanxia as his co-supervised PhD student. I highly recommend joining his Physical Vision Group (physicalvision.github.io) at NTU Singapore, and I believe it will be an inspiring place to grow as a researcher!

Thanks Zhenjun for sharing. Our PM-Loss tackles low geometry quality in feed-forward 3DGS caused by "discontinuity" in predicted depth, and features a plug-and-play training loss that leverages geometry priors from pointmap-based models like VGGT. More at aim-uofa.github.io/PMLoss

ZPressor, an architecture-agnostic module that enables existing feed-forward #3DGS models to effectively handle dense input views (up to 100 on an 80G GPU). More at: lhmd.top/zpressor/

MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views Yuedong Chen, @ChuanxiaZ, @haofeixu, @supremeZhuang, Andrea Vedaldi, Tat-Jen Cham, Jianfei Cai tl;dr: extend MVSplat to 360° NVS arxiv.org/pdf/2411.04924

Fantastic dinner @theworldlabs, glad to meet so many nice people around.

MVSplat enables super fast scene reconstructions from as few as two images in a single forward pass! Code & models: github.com/donydchen/mvsp… Joint work with amazing collaborators @donydchen @ChuanxiaZ @supremeZhuang @mapo1 @AutoVisionGroup Tat-Jen Cham and Jianfei Cai!

Well summarised, thanks for the hard work! Glad to share that me and @QianyiWu7 have managed to make our contributions to such a fast developing community. More details please refer to Sem2NeRF(donydchen.github.io/sem2nerf/) and ObjectSDF(wuqianyi.top/objectsdf/).

McDonald’s inside a park. Haven’t ate hamburgers for a long time, nice food and beautiful views make a enjoyable Saturday dinner.

After shipping a table from Taobao, my housemates start playing Mahjong every night. Interesting to find that I’m the only one in the house who don’t know how to play🧐

Finding it increasingly hard to keep up with all of the activity in deep learning right now, as # tabs -> \infty and tab width -> 0. It hasn't even been a decade since AlexNet (~Sep 2012 => ~8.5yrs) and a lot happened. Another 8.5 will be ~2030, I wonder what that's like.

#NowPlaying 500 Miles, by Peter, Paul and Mary open.spotify.com/track/6oMZhY0f…

@david2881234 What is this?

share from my ps4 pro #PS4share

@leehsienloong 7 confirmed cases today, 2 of which are taxi drivers, considering the highly contagious of the novel coronavirus, I don’t understand why govt still keeps saying the risk of infection is low. Also, 4 out of 40 cases in SG are in critical condition, looks lethal enough already.

We have faced the 2019-nCoV situation for about 2 weeks now. People are understandably anxious & fearful, but there is no need to panic — Singapore has ample supplies. Instead, let us remain united & resolute, stay calm & carry on with our lives. – LHL youtu.be/oNw1pyksKHo