Fei-Fei Li's Atlas Makes Bullet Time From Three Phones. You Can't Have It Yet.

TL;DR

World Labs, the spatial-intelligence startup Fei-Fei Li cofounded, published Atlas on September 1: a single model pretrained from scratch to read and write text, images, video and 3D, with the camera pose as a native input rather than a prayer in the prompt. It generates up to one minute of 1440p video along a camera path you draw, rebuilds a real room from two or three photos, turns three cell phones on tripods into a bullet-time rig, and outputs point clouds or Gaussian splats for robotics simulation. On World Labs' own numbers it beats five open-source 3D reconstruction specialists on average and wins 75% to 94% of head-to-head camera-following votes against five video models. It is entering early access with select partners: no weights, no parameter count, no price, no date.

What Atlas is

World Labs calls Atlas an omni world model. The architecture is a multimodal autoregressive diffusion transformer: it consumes a sequence of mixed elements (text, images, camera poses, 3D depth maps, with video represented as a sequence of images), generates each new element one at a time conditioned on everything before it, and produces each image by rectified-flow denoising. World Labs says it designed a new base architecture rather than borrowing an LLM or a video model, so it can inherit tricks from both: KV caching from the language side, diffusion distillation and classifier-free guidance from the image side.

The idea that makes the rest work is what World Labs calls a spatial context. Every image in the context is pinned to an explicit 3D camera position, so the model is not just remembering pixels, it is remembering where the camera stood when it saw them. Think of an LLM's context window as a stack of notes on a desk; Atlas's context is a pinboard with map coordinates, and every new photo you pin down tells the model where the world is still blank. That is why you can drop two unrelated reference images into the context, place them apart in 3D, and get a generated hallway that plausibly connects them.

World Labs' own line about what this buys you: "you are staging the scene, not pulling the lever of a slot machine."

Four jobs, one model

• Camera-controlled generation. From one to six reference images plus a hand-designed camera path, Atlas renders new views at any position and angle, extrapolating beyond the inputs. The showcase is a one-minute 1440p video generated from a small set of reference images.
• Spatial reconstruction. From one to dozens of photos, Atlas produces novel views and explicit 3D. World Labs says faithful reconstructions typically arrive with as few as two or three images, and that the context can hold over a hundred. The Stanford Main Quad demo uses two to twenty-five ground-level photos and then flies an aerial path over the campus.
• Space-time simulation. With footage from three to five ordinary cameras, Atlas freezes time and reframes the shot from angles no camera occupied. The same machinery reconstructs a space from a phone video and then renders what a simulated robot's body cameras would see as it moves through it.
• Image generation. Text-to-image and 360-degree panoramas, with text rendering and style range. World Labs is explicit that this is a side effect, not the point.

The camera-following benchmark

World Labs paired a single input image with one to three cinematic moves (pan, truck, crane) and asked third-party human raters which of two outputs followed the intended path. Atlas received the camera as native input; the competing video models got the same path described in text, because none of them accept camera geometry directly. World Labs acknowledges that better prompting might help the others, but text is what most people actually use.

The five opponents are Seedance 2.5 from ByteDance, FLUX 3 from Black Forest Labs, Alibaba's Happy Horse 1.1, Google's Gemini Omni Flash, and MiniMax H3. World Labs says the margin grows as the camera trajectory gets more complex, which is the expected result when one side speaks geometry and the other side is guessing what "truck left, then crane up" means.

The reconstruction benchmark, including where it loses

The more interesting table is 3D reconstruction from sparse posed views, a problem with a decade of specialist work behind it. World Labs reports mean absolute-relative pointmap error (times 10 to the minus 3, lower is better) across seven standard datasets (DTU, ETH3D, KITTI, NRGBD, 7-Scenes, Tanks and Temples, and ScanNet), and says it re-ran every baseline under one protocol rather than copying published numbers.

The baselines are serious: π³ (ICLR 2026), Pi3X run with known camera poses, VGGT-Ω 1B from the VGGT line (the original was the CVPR 2025 best paper), ByteDance's Depth Anything 3, and Meta's MapAnything. Atlas wins the average and most rows, sometimes by a lot: on ETH3D it scores 9.3 against Pi3X's 18.7, and on NRGBD 6.5 against 13.3.

Credit to World Labs for publishing the rows where it does not win. On Tanks and Temples, VGGT-Ω 1B posts 40.2 to Atlas's 42.4, and Pi3X ties Atlas exactly. On KITTI the gap to Pi3X is 60.0 versus 60.2, which is noise. A generalist that lands on top of the specialists' average while losing a couple of rows is a believable result. A generalist that swept every row would have been the suspicious one.

Bullet time from a backpack

The demo most likely to get forwarded is the reframing one. World Labs says the multiview clips were shot by its own engineers and researchers with ordinary cell phones and action cameras on tripods and clamps that fit in a backpack, three to five views per scene. Atlas reconstructs the scene from those views, after which you can freeze the moment and orbit it. The Matrix needed an arc of still cameras and a crew; this needed a few phones and someone willing to carry the clamps.

For robotics the same trick runs the other way. World Labs captured two large environments with a phone video, used 24 frames from each for reconstruction, and then rendered the RGB and depth a simulated robot's body-mounted cameras would observe along arbitrary paths. For manipulation it goes further: from a few casual recordings it helps build a simulation that captures how rigid, articulated and deformable objects move, and then lets you vary objects, positions, lighting and background to mint training data. That is the "Real-to-Sim" flywheel the company laid out in its July post on building worlds that train robots, and its July acquisition of SceniX was a bet on the same pipeline.

Where it fits in the World Labs stack

Atlas is the model underneath, not a product. World Labs says it will power future versions of Marble, the world generator it shipped on November 12, 2025, and the programmatic World API that followed on January 21, 2026. The shared currency is 3D Gaussian splats: Atlas can fill the gaps in a point cloud and emit a complete splat scene that renders on-device at high resolution and frame rate, the same representation Marble exports and that the open-source Spark renderer draws in a browser on top of three.js. If you already have a Marble-to-Spark pipeline, Atlas is a better upstream, not a new format.

The money behind it: World Labs emerged from stealth in September 2024 with $230 million and raised another $1 billion on February 18, 2026 from a group including AMD, Autodesk, Emerson Collective, Fidelity, NVIDIA and Sea, with TechCrunch reporting Autodesk's share at $200 million. World Labs has not confirmed a valuation.

What you cannot have, and what you cannot check

Atlas is entering early access with select partners; everyone else fills in a form and waits for an email. The post discloses no parameter count, no training data description beyond "a large diverse corpus of multimodal data," no inference speed, no pricing, and no timeline for the API or Marble integration. There is no paper, only a blog post with a BibTeX entry. Every number above is World Labs' own evaluation, including the baselines it re-ran, and nobody outside the company has reproduced any of it yet.

Two caveats came straight from the company on Hacker News, where a World Labs cofounder and the Atlas project lead took questions. Asked why the demos mostly freeze time while the camera moves, the cofounder said Atlas can handle some scene motion but that temporal consistency "is certainly one area we plan to improve going forward." Asked whether you could pour every geotagged photo on earth into it, the project lead said Atlas is an autoregressive diffusion model, so LLM-style context length limits apply; its edge is that the context is a sequence of posed images you can juggle creatively, not an infinite one. The cofounder also said 3D consistency as the camera moves holds even without an explicit point cloud or splat, which is the claim that matters if you want to navigate a generated world rather than just render a clip.

Why a builder should care anyway

Three reasons. First, camera pose as a native input is the right abstraction, and once one lab ships it the text-prompt-for-camera-motion era looks like a workaround. Second, the splat output plugs into an existing open pipeline (Spark, three.js, any engine that imports splats), so the day Atlas reaches the World API, Marble integrations get better without a rewrite. Third, the three-phone bullet-time and 24-frame room scan demos are things an indie VFX shop or a one-intern robotics lab could actually do, if and when the model is available to them. That last clause is doing a lot of work today.

Key Takeaways

• World Labs' Atlas is a from-scratch multimodal autoregressive diffusion transformer that treats camera pose and depth as native inputs, generating up to one minute of 1440p video with exact camera control.
• Third-party raters preferred Atlas's camera following 75% to 94% of the time against Seedance 2.5, FLUX 3, Happy Horse 1.1, Gemini Omni Flash and MiniMax H3, with rivals prompted in text.
• On sparse-view 3D reconstruction it averages 25.3 AbsRel error against 28.7 for Pi3X and 34.7 to 47.7 for the other specialists, while losing Tanks and Temples to VGGT-Ω 1B and tying on KITTI.
• Reframing works from three to five phone cameras; robotics scenes were rebuilt from 24 frames of phone video, then rendered from a simulated robot's own cameras.
• It outputs Gaussian splats, the same format Marble exports and the open-source Spark renderer draws, and will power future Marble and World API versions.
• Access is early-access with select partners: no weights, no size, no price, no date, and no independent evaluation yet.

Sources: World Labs, Atlas: A World Model for Spatial Intelligence, Hacker News discussion with World Labs staff, World Labs, funding announcement (Feb 18, 2026), TechCrunch on the $1B round, World Labs, Marble, World Labs, World API, World Labs, Building Worlds That Train Robots, World Labs on X