Pinned Tweet
Single video in, 4D human out.
4DAnyone turns a casual monocular video into a 4DGS model.
No camera rig, no calibration, no tripod.
Project: 4danyone.github.io/
4DAnyone just got easier to use!
We’ve added a GUI powered by @rerundotio. Tweak camera layouts, run inference, and explore the generated multi-view videos right in your browser.
Give it a try with your own videos!
This GUI was inspired by the 4DAnyone-Rerun Space built by @pablovelagomez1 🚀.
1
5
400
Sooo cool! Thank @mickmumpitz for trying out 4DAnyone
2
692
Happy to release 4DAnyone-Turbo, a faster and lighter model that:
• runs with just 22 GB of peak VRAM
• generates a 121-frame video in 27 seconds on a single RTX 4090
• turns a monocular video into synchronized, view-consistent videos
Try it on a consumer GPU:
github.com/ant-research/4DAn…
Special thanks to @pablovelagomez1 for his experiments and guidance on speeding up 4DAnyone!
3
804
Cooool results! Makes me hopeful we can integrate an open-source 4DGS pipeline into 4DAnyone
My latest 4DGS output using 4DAnyone -> 4D-GS. It's not the BEST output, but I also used a 2 year old 4DGS pipeline that I feel may be the limiting factor here.
Keep in mind, the source data was a single 2D video, so its amazing its 4D in the end!
Processing time for this 4 second scene: 69 minutes on a RTX 6000 Ada. Not realistic to do high resolution long videos on a workstation.
#3D #4DGS #Computervision
1
9
1,075
Amazing results! Love seeing 4DAnyone being used in your pipeline 🚀 Really excited to see where you take this next!
Researching DIY dynamic gaussian splats ("4DGS") captured w/ just 1-4 cameras & played back with @theworldlabs @sparkjsdev + @threejs in real-time in the browser
It's underexplored territory, with a few hard problems re: processing, compression & streaming. So far I've spun up cloud H100s, generated enough novel views to train gaussian frames, and gestured wildly to camera as if in a silent film. This is mostly 4DAnyone & SharpML, but I also tried Lift4D and StudioRecon
Already these crude prototypes show a fluidity that is often missing from mocap-to-avatar 3D representations of humans. My goal is affordable capture of a wider emotional and physical range of human performance for games and entertainment than we can achieve with keyframe-driven skeleton rigs. (Is realistic 3D human representation not also useful in the realm of robotics sim & training?)
To that end, this tech needs to escape the realm of expensive 32- and 64-plus-camera capture studios. These tests rely on as few as one iPhone, with dense views generated by a variety of models and techniques on NVIDIA hardware and open models like GEN3C and Cosmos via Google Cloud (cc: @realmattbcool & @asierarranz re: #NVIDIAGTC!)
@RadianceFields reported that MPEG is seeking training data for a 4DGS compression standard - I'll submit some of these clips since format and streaming are the current bottlenecks for web production. I wonder how long that process will take?
Just scratching the surface here, but with new world models like Atlas coming out, I'm eager to keep investigating! Hmu if you're thinking about the same 🤙 More soon
4
494
Yudong Jin retweeted
Alright, I went ahead and built 4DAnyone into a proper @Gradio space. I thought it was cool enough that it deserved the effort and made sure to take advantage of our @rerundotio Gradio integration.
TLDR takes in a single video and is able to output multiview consistent videos. Its kinda incredible how good the quality is. Got it working on my 5090 + DGX Spark, so it works with <32GB cards =]
Sadly, I had to make this version only output 6 videos; otherwise, it would take way too long, but its able to go all the way up to 48 videos locally. Just takes like ~20-30 minutes.
No splatting on this one, as again it would just take too much GPU time that ZeroGPU won't easily support and will eat up all your quota
This also feels like the perfect fit for a new Gradio Workflow, something I'll look into at some point.
Inference time ~5-8 minutes on the space, sped things up in the video to keep it digestible
I got nerdsnipped by @krahets with 4D Anyone. I was working on part 3 of the arkitscenes blog post for @rerundotio, but I just couldn't help but try to bring in 4d splatting from a single video into rerun since we shipped splatting support.
Did a few things here
1. spent a bunch of time optimizing the DiT model so it would run on my 5090. Got memory from ~50gb -> less than 32gb. No need for enterprise cards or super slow dgx spark
2. more importantly, I got the runtime down to ~5 minutes for going from 1 video -> 16 views. This was nearly 40 minutes before on my 5090!
3. I also got a version of 4d splatting training working, its a mess and doesn't look great but hey it works!
Really really cool work, hoping to do more with it in the future when I'm a bit less scattered with all the cool things I want to work on 😅
Awesome work! Thanks, Pablo! I’m also working on cutting 4DAnyone’s peak VRAM with better memory management and speeding it up with TensorRT and distillation. Hopefully we can combine our improvements and make it much easier to run. Can’t wait to see what you do next! 🚀🚀
I got nerdsnipped by @krahets with 4D Anyone. I was working on part 3 of the arkitscenes blog post for @rerundotio, but I just couldn't help but try to bring in 4d splatting from a single video into rerun since we shipped splatting support.
Did a few things here
1. spent a bunch of time optimizing the DiT model so it would run on my 5090. Got memory from ~50gb -> less than 32gb. No need for enterprise cards or super slow dgx spark
2. more importantly, I got the runtime down to ~5 minutes for going from 1 video -> 16 views. This was nearly 40 minutes before on my 5090!
3. I also got a version of 4d splatting training working, its a mess and doesn't look great but hey it works!
Really really cool work, hoping to do more with it in the future when I'm a bit less scattered with all the cool things I want to work on 😅
Happy to share that we’ve reduced 4DAnyone’s peak VRAM while slightly improving inference speed. A 6-view group now peaks at around 25 GB, allowing full 24-view generation on a 32 GB GPU. Details: github.com/ant-research/4DAn…
3
216
Paper: arxiv.org/abs/2608.20335
Code: github.com/ant-research/4DAn…
Model: huggingface.co/AntResearch/4…
Yudong Jin retweeted
🚀 LingBot-VA 2.0 is here!
After half a year of teamwork, we're thrilled to release LingBot-VA 2.0 — a native video-action foundation model for generalizable robot control.
Unlike prior world-action models that retrofit generic video generators for robot control, LingBot-VA 2.0 is natively pretrained from scratch as a video-action foundation model.
Three key insights:
🌍 Native Video-Action Pretraining for learning world knowledge that enables strong generalization.
🧩 Semantic Visual-Action Tokenizer for more accurate action prediction with robust prompt following
⚡ Foresight Reasoning enables the robot to think ahead while acting, delivering continuous, responsive control without interrupting execution.
It runs in real time on consumer-grade GPUs, supports up to 150 Hz control, and generalizes to unseen tasks.
👇 Demos below
Want to reconstruct very large-scale scenes accurately?
Check out our Scal3R (CVPR'26 Highlight) — scaling up VGGT with TTT Layers: Split image sequences into chunks, compute TTT gradients per chunk, and then average gradients across chunks.
Code is out: zju3dv.github.io/scal3r/.
Yudong Jin retweeted
🔥 Very excited to share that we’re releasing LingBot-World 🌍 @robbyant_brain — an open-source frontier world model!
We’re pushing the limits of:
🔹 High-Fidelity Simulation & Precise Control
🔹 Long-Horizon Consistency & Memory
🔹 Modeling Physical & Game Worlds
The most surprising part? The emergence of sophisticated behaviors that go beyond simple video generation.
👇I’m obsessed with this dragon demo 🐉. It can rollout for 1 min while maintaining crisp visual dynamics and consistent memory!








