Weijie Wang

🚀 Introducing World-R1: Video models already know 3D — they just need RL to wake it up! No arch changes. No video training data. No extra inference cost.⬇️ 🌐Website: aka.ms/world-r1

@yourkaisensei You only need to design custom text prompts and trajectory generation strategies. We aim to offer a new perspective instead of merely a trained model.🥰

@wjwang2003 curious how you’re checking for reward hacking here. do the 3D gains hold on weird camera paths outside the training trajectory distribution?

🚀 Introducing World-R1: Video models already know 3D — they just need RL to wake it up! No arch changes. No video training data. No extra inference cost.⬇️ 🌐Website: aka.ms/world-r1

@yourkaisensei Most camera gains come from latent injection. You can add randomization in training to adapt to diverse camera paths.

@MrManderly It’s just because of X’s video size limit. Not heavy computation. I’m not sure if upgrading helps🤣

@wjwang2003 This is fantastic work. Can we infer that this is currently very expensive computationally from the low frame rate examples?

World-R1 has been accepted by @icmlconf , see you in Seoul! 📄 Paper: arxiv.org/abs/2604.24764 📷 Code: github.com/microsoft/Worl…

@aswinrrv @AzmineWasi @icmlconf someone say they have another type

@AzmineWasi @icmlconf Can anyone confirm if everyone is seeing this, even with higher scores?

@icmlconf ICML Position Paper decisions seems out, indirectly 👀 Public-release or In-person presentation...?

@AzmineWasi @icmlconf I have the same question

@taziku_co Thanks for sharing @taziku_co !

World-R1 Reinforcing 3D Constraints for Text-to-Video Generation microsoft.github.io/World-R1/ via:@wjwang2003

動画モデルは、最初から3Dを少し知っていた！？。 World-R1は、既存のText-to-Videoモデルに対して、追加推論コストなしで3D整合性へ寄せ、物体の永続性・幾何一貫性・カメラ制御を改善する方法。 人間評価でも幾何一貫性92%と非常に高いスコアを記録 詳細は🧵

Microsoft presents World-R1 Reinforcing 3D Constraints for Text-to-Video Generation paper: huggingface.co/papers/2604.24…

@_akhaliq Thanks for sharing @_akhaliq! 🙏 🏠 Project: aka.ms/world-r1 💻 Code: github.com/microsoft/Worl…

@newlinedotco Totally agree on the inference tax concern! Good news: World-R1 adds zero overhead at inference. 3D foundation models and VLM critics only serve as reward signals during RL training. Once trained, it's the same architecture, same speed, just better 3D understanding built in ✅

@wjwang2003 video models knowing 3d is a massive unlock but the real hurdle for agents has always been the inference tax of running these heavy vision-to-action loops.

Microsoft just released World-R1 A framework that aligns text-to-video generation with 3D constraints through reinforcement learning, using feedback from pre-trained 3D foundation models to enforce structural coherence without altering the underlying architecture.

@wjwang2003 , Xiaoxuan He, Youping Gu, @Yif_Yang, @SteveZeyuZhang , Yefei He, Yanbo Ding, Xirui Hu, @donydchen , Zhiyuan He, Yuqing Yang, @supremeZhuang

📄 Paper: huggingface.co/papers/2604.24… 🏠 Project: aka.ms/world-r1 💻 Code: github.com/microsoft/Worl…

🔑 How it works: • Embed camera trajectories into diffusion noise, zero extra modules • 3D rewards from Depth Anything 3 + Qwen3-VL as geometry critics • Periodic decoupled training: buildings stay rigid, flags still wave 🏗️🚩 • 3K text prompts only, no video data

Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective @wjwang2003, Qihang Cao, Sensen Gao, @donydchen, @haofeixu, @wenjing_bian, @songyoupeng, Tat-Jen Cham, @ChuanxiaZ, Andreas Geiger, Jianfei Cai, Jia-Wang Bian, @supremeZhuang tl;dr: new survey arxiv.org/abs/2604.14025

@zhenjun_zhao @donydchen @haofeixu @wenjing_bian @songyoupeng @ChuanxiaZ @supremeZhuang Thanks @zhenjun_zhao ! Check out our paper list and project page here: 📷 Github: github.com/ziplab/Awesome… 🌐 ff3d-survey.github.io

Authors: @wjwang2003 , Qihang Cao, Sensen Gao, @donydchen , @haofeixu , @wenjing_bian , @songyoupeng , Tat-Jen Cham, @ChuanxiaZ , Andreas Geiger, Jianfei Cai, @jiawangbian, @supremeZhuang

📢 We release "Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective" — a comprehensive survey covering 200+ papers on feed-forward 3D reconstruction! Instead of categorizing by 3D representations, we propose a problem-driven taxonomy. 🌐 ff3d-survey.github.io

A joint effort by ZJU, NTU, Monash, ETH Zurich & Uni Tübingen. 📄 Paper: huggingface.co/papers/2604.14… 💻 GitHub: github.com/ziplab/Awesome… Feedback & discussions welcome! 🙌

6+ application areas (AD, robotics, SLAM, video gen…)