Wei Yu

Dropping an exciting new demo of MosaicMem! 👀🔥 A friend brought up a great question: why not combine long-horizon navigation video generation, promptable world events, and scene concatenation? Fair point — so we gave it a shot. 🎬✨ For more technical details, check this thread 🧵👇 x.com/GnosisYu/statu… #WorldModel #GenerativeAI #VideoGeneration #InteractiveAI #Genie3 #EmbodiedAI #GameAI

📢📢📢 Introducing "FullCircle: Effortless 3D Reconstruction from Casual 360° Captures" TL;DR: 10x faster casual capture with clean reconstructions Homepage: theialab.github.io/fullcircle Code: github.com/theialab/fullc… arXiv: arxiv.org/abs/2603.22572 Led by Yalda Foroutan & Ipek Oztas

🚀 Check out our new demo using MosaicMem! It combines long-horizon navigation, promotable events and multi-scene concatenation 🌏

Matrix-Game 3.0🔥real-time interactive world models from @Skywork_ai huggingface.co/Skywork/Matrix… ✨ MIT license ✨ 720p @ 40FPS with a 5B model ✨ Minute-long memory consistency ✨ Unreal + AAA + real-world data ✨ Scales up to 28B MoE

Uni-v1 is finally here!! A state of the art model that outperforms Nano Banana / GPT along many axes that you can now try for free 👀 🥰 Building this model brick by brick from scratch was the highlight of my life ❤️ model is fully available now! We are COOKING rn @LumaLabsAI

Introducing EgoVerse: an ecosystem for robot learning from egocentric human data. Built and tested by 4 research labs + 3 industry partners, EgoVerse enables both science and scaling 1300+ hrs, 240 scenes, 2000+ tasks, and growing Dataset design, findings, and ecosystem 🧵

8 years ago, it was just a toy. All things start as a toy, when they don’t have a name or feel strange, that’s where the most interesting things are happening.

Adding a few tags for visibility 👇 #Genie3 #WorldModels #VideoGeneration #AI #Game

World models have made impressive progress in video generation, yet they still struggle with a fundamental challenge: memory. In long rollouts, the camera trajectory gradually drifts from the user-specified motion and revisited scenes no longer align with earlier observations. These errors accumulate over time, causing the generated world to steadily lose coherence. 🚀Excited to share our solution MosaicMem 🌍🧠 — our new hybrid spatial memory for video world models. Project Page: mosaicmem.github.io/mosaicmem/ Paper: huggingface.co/papers/2603.17…

Huge thanks to all the amazing collaborators on this work 🙏 @Runjia Qian @MqquQ0nIwoXrU4M @Yumeng Li @YumengLi_007 @Liquan Wang @LiCHO13676030 @Songheng Yin @zha_bu_duo_deLe Sri Siddarth Chakaravarthy P @Dennis Anthony @DennisAnthony__ @Yang Ye @YangYe20161101 Yidi Li Weiwei Wan @Animesh Garg @animesh_garg Couldn’t have done this without you all — truly appreciate the effort and collaboration! 🚀

@HuggingPapers Thanks so much @HuggingPapers for sharing our work! 🙏 We’ve just updated the post today with a more detailed demo video — would love for you to check it out! x.com/GnosisYu/statu…

MosaicMem A hybrid spatial memory for video world models bridging explicit 3D and implicit memory, enabling long-horizon navigation, memory-based editing, and dynamic scene generation with improved camera consistency.

@kwangmoo_yi Huge thanks @kwangmoo_yi for the share!! 🙏 We just updated the post with a more detailed demo video today — would really appreciate you taking a look! x.com/GnosisYu/statu…

Yu et al., "MosaicMem: Hybrid Spatial Memory for Controllable Video World Models" A patch-based spatial memory that you raster into views + glues to make things work.

A key strength of MosaicMem lies in enabling dynamic, evolving scenes 🎮🌍. It supports promptable world events while preserving a persistent background, effectively combining controllable dynamics with reliable camera control. In addition, it enables memory manipulation for scene editing—composing elements from different scenes into a unified generation, creating composite experiences that do not exist in reality while maintaining spatial coherence.

The core problem is memory. However, the puzzling part is that today’s memory designs perform well in some regimes yet fail in others, motivating a closer examination of prevailing paradigms and their limitations. - explicit 3D memory (reprojection-based): geometrically solid, but fails to preserve the prompt-following capability of pretrained models. - implicit latent memory: handles dynamics fine, but camera drift creeps in and compounds Instead of posing this as a tradeoff, MosaicMem combines the best of both worlds.

@HuggingPapers Thanks @HuggingPapers for sharing! I’ll be posting a more detailed breakdown of our Mosaic Memory tomorrow—stay tuned 👀

@lily_goli Same here! This was probably the best reviewing experience I’ve had so far — lots of genuinely enjoyable papers.

wow I guess I was lucky with this year's CVPR reviews, more than half of mine were actually nice papers I enjoyed reading.