Introducing Atlas: The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D. Model the world, move the camera, and simulate space & time.

We pre-trained Atlas from scratch to take multimodal inputs, including camera movement, and turn it into 3D grounded views. Atlas puts you in control: direct the views, reconstruct real spaces with your inputs, and build explorable worlds. Read more:

By positioning multiple input views within the model's spatial context, Atlas allows you to generate image and video frames with precise control. Here, we hand-designed a camera trajectory to generate a 1 minute video at 1440p resolution from seven reference images.

Atlas also outputs explicit 3D from one to many input images, beating top open source reconstruction models. Passing more images gives Atlas more context: the more it sees, the less it imagines.

Atlas understands both the spatial structure of the world and how it evolves over time. This lets us turn a handful of ordinary cameras into a "bullet time" multiview capture studio, capturing dynamic moments frozen in time.

For robotic simulation, Atlas reconstructs a space from just a few photos and generates the photorealistic RGB and depth data any robot's sensors would observe on any trajectory. Robots can now be trained and tested in far more spaces. Until now, scanning spaces like these required expensive equipment and time-consuming capture.

All of this is powered by one unified architecture: a multimodal autoregressive diffusion transformer, pretrained from scratch. This foundation blends the best of modern LLMs and video models, benefiting from the architectural, algorithmic, and systems advances from both areas.

Atlas is our scalable foundation model that can perceive, generate, reason, and interact with virtual and physical worlds. We're opening up early access in the coming weeks. Learn more and sign up here:

this is like the millionth ai thing called atlas. pls come up with a different name.

Amazing work! Great step towards real2sim for robotics!

This looks like it could have so many uses in games and movies. And VR!

the amount of slop it enable is unimaginable - but seriously world models are also in this everyday release cadence from today

This looks pretty great, congrats to the whole team!

with this, so many unlocks, reducing the gap with sim-2-real, video production, modeling, visual simulations, the use-cases are almost limitless. unbelievable work, and this is just the start.

The video quality is looking fabulous. Much better than any models..

turns out s-tier research talent, lots of compute, and a bold vision can produce outstanding models

Give me a SLOP copy right there.

this looks amazing, wow

Atlas (spatial context) + an approach like Genie 3 may be a way to get persistent worlds that you can interact with. 3d information as a modality in and out from multimodal models.

generated 3D worlds have to remain fast to retrieve across repeated simulation cycles

I knew it. Pokémon Go was to train this and make a 3d model of the whole world so AI can use the datacenters to make world sims and train AGI

👏👏👏 🖖

Looks super realistic!

this has me more excited than fable 5.1

incredible, congrats!

@levelsio With the speed of AI these days, I don’t understand the blog about a model being released instead of access to the model. We have 6 other new toys we can play with per day, we’ve already forgotten about your dumb blog

Incredible!

Congrats @drfeifei and team eager to test it!

What pixel perfect means in this case? That there are no seams and bugs - because they are artificially adapted, or that I can scan an area with 3-4 cams and take measurements?

stunningly beautiful. excited to see new worlds unlocked!

people have no idea on how big of a deal this is.

@martin_casado let’s go, this is huge!

Go @theworldlabs ! Go Atlas! What a dramatic improvement from our initial Marble launch a year ago!

World Labs but I need VPN to access it from Lagos, Nigeria. Shame.

camera-path control is easy to fake with a good video model. actual 3d reconstruction underneath it is the part that separates a real world model from a video model wearing a trenchcoat. curious how it holds up on occlusion once the camera swings past 90°

This is insane

Bullet time really puts a world model’s grasp of space and time to the test: a few ordinary phones can freeze a moment, then let you orbit it from any angle with no dead spots, geometrically consistent, and even continue the motion into subsequent animation. It turns “filming” into directing inside a world the model already understands. Stunning.

I wanted to see if AI could do more than generate a beautiful image. So I tested World Labs’ Marble with a pretty demanding prompt: An abandoned luxury railway station in the mountains — multiple rooms, platforms, stairs, furniture, windows, railway tracks, mountains, rain, lighting and realistic spatial relationships. Then I just started moving around. 👀 And honestly… this surprised me. It’s definitely not perfect. Some textures get blurry, some geometry breaks, and you can still spot the AI artifacts. But the interesting part is that I’m not looking at a single generated image. I’m actually exploring a generated 3D world. 🌍 We’ve spent the last few years asking AI: “Can you generate what this looks like?” Now the question is becoming: “Can you generate the place itself?” That shift feels pretty significant. 🚀 Still early days, but this space is getting VERY interesting.

Looks so realistic!

This is way too impressive holy shit

can't wait!