Teardown · World model · 2026-09-21

Atlas, taken apart

World Labs released Atlas as a blog post: no paper, no code, no evaluation artifacts. This page walks the whole machine anyway, from what a world model is supposed to do to where this one stops, and it marks every claim with where the claim came from.

Provenance All three shown
01

What a world model is asked to do

Fei-Fei Li's framing sets three requirements: a world model should be generative, producing worlds that are perceptually, geometrically and physically consistent; multimodal, taking images, text, gestures and actions; and interactive, outputting "the next states based on input actions". WL

World Labs' own taxonomy sorts models by what they emit inside the agent loop. Pick a class to see what it produces and who sits there.

Functional taxonomy · agent → action → state → observation
World Labs calls the simulator "the linchpin".

Where Atlas actually sits. Its published inputs are text, images, camera poses and depth maps. There is no action input and no force input, so a camera pose moves a disembodied viewpoint rather than acting on the world. As disclosed, Atlas is a geometry-grounded renderer plus a state estimator, not a dynamics simulator, and on the interactive axis it sits behind action-conditioned models such as Genie 3. inf

02

Why the two obvious approaches fall short

The honest version of this argument is not that language models and video models fail. It is that neither carries a structural guarantee of physical fidelity: nothing in either objective is violated by a scene that cannot exist.

Language models

State is a one-dimensional token sequence with no SE(3) structure and no metric grounding, so spatial relations have to be rebuilt out of linguistic regularity. Li cites near-chance distance estimation. This is an inductive-bias mismatch, not a scale problem. inf

Video diffusion

The model learns p(x1:T | c) with an unobserved camera, so ego-motion and scene content stay entangled. There is no global frame, and memory is a bounded temporal window: a place revisited after it leaves the window is regenerated rather than recalled. Geometry is emergent, and dynamics learned from passive video leave p(st+1 | st, at) unidentifiable without action labels. inf

The counter-evidence, kept in view. Large video models do show emergent 3D consistency, with outputs that are often reconstructible by structure from motion, and Wiedemer et al. (2025) reported zero-shot physical reasoning in Veo 3. The defensible distinction for Atlas is narrower: explicit pose conditioning removes one source of drift. It says nothing about dynamics. inf

03

Four words, one machine

World Labs describes Atlas as a multimodal autoregressive diffusion transformer. Each word carries a specific commitment. WL

Text, images, camera poses and depth maps, with video treated as a sequence of images. Every image and depth map is conditioned on an explicit camera pose, which is the whole trick: geometry is an input type, not something the model has to infer from pixels. WL

The unit is an element, a whole image or depth map or pose, not a token. Each output is generated one at a time, conditioned on what came before, so a task is a sequence of inputs followed by outputs. Within an element, generation is diffusion; across elements, it is autoregressive. That places Atlas beside MAR, Transfusion and Diffusion Forcing rather than beside a plain video diffusion model. WL

A rectified-flow latent diffusion model, with distillation, classifier-free guidance, shifted noise schedules and VAE improvements named as compatible techniques. Section 04 makes the flow part interactive. WL

Which buys the serving stack that language models already have: KV caching, cache-aware routing, disaggregated serving. That detail is also a tell about the attention mask, below. WL

What KV caching implies about attention
A cache is only valid if earlier elements never attend to later ones and the cached context stays fixed. That implies a block-causal mask: bidirectional inside the element being generated, causal to everything before it. World Labs has not published the mask. inf
04

Rectified flow, by hand

A diffusion model has to carry noise to data. Rectified flow does it along paths that are as close to straight as training can make them, which is why so few integration steps are needed and why the model distils well. Drag the slider: every particle travels its own straight line from a noise sample to a data point.

z(t) = (1 − t)·zdata + t·ε   |   target velocity = ε − zdata   |   loss = ‖ vθ(z(t), t | context) − (ε − zdata) ‖²
Noise → data, K steps t = 1.00 · pure noise
K = 6
Straight paths mean a coarse step lands near the true trajectory, so K can be small. Curved paths, as in a classical diffusion schedule, need many more. The shifted timestep schedule World Labs mentions biases training toward the high-noise end, which matters at 1440p because a higher resolution keeps more signal at the same t. WL
05

Spatial context, and why it has to be managed

Every image is grounded at a 3D position, and those posed images form the spatial context. The context holds more than 100 images, and two unrelated photographs placed in 3D produce a world that interpolates between them. Memory is organised by where, not by when. WL

How poses are encoded is not published. The two live candidates are Plücker ray maps, which attach a six-vector per pixel and need a canonical world frame, and relative projective attention, which injects the relative transform between two views into the attention logits and stays invariant to the choice of world frame. Pixel-perfect control over 100+ views favours the relative family, but the evidence cannot settle it. inf

Not "attention in 3D Euclidean space". Dot-product attention is not distance-aware by itself. Geometry enters through the pose encoding. A literal 3D mechanism, rotary embeddings over unprojected coordinates or k-nearest-neighbour attention, is possible but unsupported by anything published. spec

Token budget for the one-minute demo
24 fps 60 s

Assuming 8× spatial compression, 2×2 patches and 4× temporal compression. The bar is scaled against a one-million-token reference. A single forward pass over the full minute is not on the table, so the video has to be produced in chunks from a selected context, which is what "spatial context management" has to mean. inf

What holds a minute together

  1. Exogenous poses. The trajectory is given rather than predicted, so pose error cannot accumulate. This is the structural advantage over text-prompted video.
  2. Persistent anchors. Reference images stay in context instead of ageing out of a temporal window.
  3. Spatial retrieval. Context selected by co-visibility rather than recency, so a revisit conditions on frames that saw the same surface.
  4. Joint depth. Generating depth alongside RGB pushes the output toward geometric self-consistency.

"No drift" is unproven. World Labs showed one hand-designed one-minute 1440p path and several paths through one scene. No revisit PSNR, no reprojection or loop-closure error, no MEt3R. The expected failure modes are retrieval misses on revisit, hallucinated content hardening once a generated frame becomes context, scale ambiguity from a single reference image, and dynamic content, for which static-scene consistency says nothing. A controlled revisit test would settle it in an afternoon. inf

06

End to end, one stage at a time

This is the whole path from a handful of phone photographs to a scene a robot can be driven through. Step through it; each stage says what is disclosed and what is a reasonable guess.

Pipeline
07

Does the evidence carry the claims?

Camera-controlled generation

Human raters preferred Atlas over MiniMax H3 (75%), Gemini Omni Flash (81%), Happy Horse 1.1 (86%), FLUX 3 (93%) and Seedance 2.5 (94%). WL

What that comparison actually measures. Atlas received native camera paths; every baseline received a text description of the path, which World Labs acknowledges. The result supports "native pose input beats prompting for camera control". It does not support "Atlas is the better generator". Rater counts and confidence intervals are not reported. inf

3D reconstruction

Mean pointmap AbsRel error, ×10⁻³, lower is better. Click a column to sort. WL

Pointmap error by datasetsorted by Avg
Method Avg DTU ETH3D KITTI NRGBD 7-Scenes T&T ScanNet
Atlas's seven per-dataset values average to exactly 25.3, which confirms the numbers were read correctly from the page.

Protocol gaps that matter. Atlas receives images and their camera poses. Only Pi3X is labelled posed; if the others ran unposed they solved a harder problem, and that is undisclosed. Against the one confirmed posed baseline the margin is modest: 25.3 against 28.7, a tie on Tanks and Temples, 60.0 against 60.2 on KITTI. Views per trial, alignment, seeds and variance are unreported, there is no third-party reproduction, and one outlet reports a prior contamination warning for VGGT-Ω unverified. inf

What the evidence does support: native pose conditioning beats text-prompted camera control, and Atlas is competitive to modestly better at posed sparse-view pointmap prediction.

08

The robotics claim, measured

As a simulated robot moves, Atlas generates the RGB and depth its sensors would observe. Two large environments were reconstructed from 24 phone-video frames each, and views were generated from body-mounted cameras. WL

The simulator supplies the base pose; forward kinematics and the camera mount give Twcam = Twbase · Tbaselink(q) · Tlinkcam, which becomes an Atlas query pose with matched intrinsics. inf

Two ways to use it, and only one is practical at training scale
 Render from the Atlas-built splatQuery Atlas per frame
Latencymilliseconds, real timeseconds, K flow steps per chunk
Determinismdeterministicstochastic unless the seed is pinned
Unseen regionsholes or blurplausible hallucination
Dynamicsstaticstatic, there is no action input
The post does not say which mode the navigation demos use. For training at scale only the first is practical today. inf

Six limits on closed-loop policy training

Who does the physics. Atlas "aids in building a simulation"; the articulated and deformable behaviour in the demos is credited to the real-to-sim workflow, not to Atlas. The likely division is appearance and geometry from Atlas, dynamics from a physics layer with system-identified mass, friction and joint axes, validated by matched open-loop replay of identical action sequences in simulation and reality. inf

09

The vocabulary, with something to hold on to

Every card has a plain definition and, underneath, the analogy that makes it stick. Click to flip.

How the pieces combine. Bricks, the autoregressive elements, are laid on a surveyor's map, the spatial context. Each brick is driven out of the fog by rectified flow, its placement fixed by six numbers and its shape read off a pin bed of depth. The finished wall is converted into soap bubbles, the Gaussian splat, for real-time viewing.

10

What I would want next

The gap between this and an interactive simulator is one input: actions. Everything else in the spatial-context design, posed elements and depth as a first-class modality, is already close to geometry-aware policy work.

  • A technical report with the pose protocol stated for every reconstruction baseline.
  • A long-horizon revisit metric, so "consistent for a minute" becomes a number rather than a demo.
  • Inference throughput, which decides whether generative querying can ever sit inside a training loop.
  • A multi-task sim-to-real rank correlation, not one task at one checkpoint.
  • A plain splat reconstruction from the same 24 frames as a baseline, to isolate what generative gap-filling adds.

The testable question underneath all of it: does open-loop generative fidelity predict closed-loop policy success? Report the two as separate axes, with three or more seeds, and the answer is worth a paper on its own.