Atlas: A World Model for Spatial Intelligence

TL;DR

Atlas is World Labs’ next-generation world model — a multimodal autoregressive diffusion transformer pretrained from scratch on text, images, video, and 3D that unifies four tasks under one architecture: camera-controlled generation (up to 1-minute video at 1440p from one or more input images with pixel-perfect trajectory control), sparse-view 3D reconstruction (image frames + point clouds + 3D Gaussian splats, reportedly outperforming specialized open-source reconstruction models), space-time simulation (video reframing / “bullet time” from 3–5 ordinary cameras, plus a Real-to-Sim workflow for robotics), and text-to-image generation (including 360° panoramas). The organizing primitive is a spatial context — each input image and depth map is grounded at an explicit 6-DoF camera pose, so the AR model conditions on and predicts elements within a shared 3D coordinate frame. Atlas is billed as the foundation model that will power future versions of Marble and other World Labs products; it enters early access with select partners at launch.

Key claims

Atlas is pretrained from scratch as a multimodal autoregressive diffusion transformer — AR over a sequence of multimodal elements (text tokens, images, camera poses, depth maps) where each element is generated diffusively by rectified flow one at a time [§Technical Details / Model Architecture].
Each input image and depth map is conditioned on an explicit camera pose, forming a spatial context rather than a purely temporal context — the AR sequence lives in 3D coordinates, not just tokens-per-step [§Technical Details / Model Architecture; §Camera-Controlled Generation / Generating with Spatial Context].
Video generation supports up to 1-minute 1440p from a small number of reference images with hand-designed camera trajectories, driven by a native camera-geometry input (not a text-prompt approximation) [§Camera-Controlled Generation / Controllable Long Videos].
Sparse-view 3D reconstruction outperforms state-of-the-art specialists on standard benchmarks despite being an omni-model; the reported protocol reproduces baselines under a shared evaluation and predicts one 3D point per input pixel [§Technical Details / Benchmarks].
On camera-conditioned generation, Atlas beats recent video models under a third-party human-rater study over cinematic camera-motion prompts (pan / truck / crane, etc.); the advantage widens as trajectories grow more complex [§Technical Details / Benchmarks].
Explicit 3D outputs (point clouds → 3D Gaussian splats) come from joint generation of new views and depth prediction; Atlas fills in regions no camera ever saw, matching the representation Marble uses so the same model integrates with the rest of the World Labs product line [§Spatial Reconstruction / Explicit 3D Outputs].
Compose-two-unrelated-images-into-one-world: placing two input images at chosen 3D positions inside the spatial context produces a coherent scene with model-imagined transitions (doorways, hallways, nooks) between them [§Camera-Controlled Generation / Generating with Spatial Context].
Robotics real-to-sim: two large environments were reconstructed from ~24 cell-phone-video frames each and used to simulate robot navigation, with Atlas generating body-mounted camera RGB + depth along arbitrary trajectories from within the reconstructed scene [§Space-Time Simulation / Robotics Simulation].
Space-time reframing (“bullet time” from 3–5 cell-phone tripods) works from ordinary consumer capture equipment, not a specialized multiview rig [§Space-Time Simulation / Reframing Video].
Scaling behavior: World Labs trained a series of models of increasing size and training compute, and reports each new compute level “unlocked new model capabilities,” with the expectation that the trend continues [§Technical Details / Model Scaling].
Atlas is positioned as the architectural bet for a unified world model — architecturally departing from both standard LLMs and standard video models to make spatial control central rather than bolted on [§Technical Details / Model Architecture].

Method

Atlas is a decoder-style autoregressive transformer whose sequence elements can be any of four data types: text, image, camera pose, and 3D depth map (video is treated as a sequence of images). Each image and depth map is bound to an explicit camera pose in a shared 3D frame — this is the “spatial context.” At generation time, the model produces the next element conditioned on all earlier elements in that sequence, and for the continuous element types (images, depth) the generation is a diffusion process — specifically a rectified flow that gradually denoises the output. The transformer backbone is described as “well-adapted to modern hardware,” and the design explicitly inherits serving primitives from the LLM stack (KV-caching, cache-aware routing, disaggregated serving) and training/algorithmic primitives from the video model stack (diffusion distillation, classifier-free guidance, shifted noise schedules, VAE design). The result is a single unified architecture that can be steered by placing images at different 3D positions in the spatial context and asking for the “next” image at any target camera pose.

Four task surfaces are exposed on the same model:

Camera-controlled generation — one or more input images plus a target camera trajectory; outputs image or video frames at the requested viewpoints, up to 1 minute at 1440p.
Spatial reconstruction — 1–100+ input images with poses; outputs novel views and explicit 3D as point clouds or 3D Gaussian splats produced by jointly generating RGB + depth. Faithful reconstruction is claimed with as few as 2–3 images; more input images monotonically reduce the fraction that Atlas imagines vs. reproduces.
Space-time simulation — from ~3–5 ordinary cell-phone-on-tripod cameras, freeze time and reframe from novel viewpoints; or reconstruct a scene from a short cell-phone video and roll out an embodied camera trajectory with matching RGB and depth for robotics.
Image generation — text-to-image and text-to-360° panorama, treated as the marginal task on the same model.

Benchmarking uses human raters for camera-conditioned generation, comparing Atlas’s native camera-pose input against baseline video models given cinematic-camera-term text prompts. The 3D-reconstruction evaluation reproduces baseline specialists under a shared protocol on multiple state-of-the-art sparse-view benchmarks. No parameter count, training data description, training compute figure, VAE choice, or quantitative benchmark numbers are disclosed in the launch post.

Results

The post reports qualitative-plus-relative evidence rather than tabulated numbers. On camera-controlled generation, Atlas beats a “selection of top-performing video models” in a third-party human-rater A/B on cinematic camera-motion adherence, with the gap widening on more complex multi-motion trajectories [§Technical Details / Benchmarks]. On sparse-view 3D reconstruction, Atlas is reported to outperform “the best specialized open-source reconstruction models” on several state-of-the-art benchmarks under a shared protocol reproducing all baselines. On robotics real-to-sim, two large environments were captured with ~24 frames of cell-phone video each and used to simulate different robot embodiments navigating different paths with matching RGB + depth from body-mounted-camera perspectives — presented as videos rather than metrics. On space-time reframing, “bullet-time” shots were generated from 3–5 cell-phones-on-tripods captures without professional equipment. The scaling section reports monotone capability unlocks across “a series of models of increasing size and training compute” during development, without disclosing the axis or the delta magnitudes.

Why it’s interesting

Atlas is the first public World Labs artifact that concretely instantiates the “renderer + simulator” unification argued for in A Functional Taxonomy of World Models — one architecture that emits both pixel views (renderer output channel) and explicit 3D state (simulator output channel), with camera pose as a first-class input rather than a text approximation. Where Building Worlds That Train Robots — Real-to-Sim-to-Real (R2S2R) as a Scalable Engine for Training and Evaluating Robot Policies operationalised the simulator-as-linchpin claim behind a proprietary R2S2R engine, Atlas is the load-bearing model underneath — and the post-hoc claim is that the split between rendering and simulation was always the wrong factoring. The architectural bet — spatial context as the AR sequence type — is a sharp contrast with the “frozen Wan2.x + adapter” recipe that dominates the current Camera-Controlled Video Diffusion cluster (VerseCrafter: Dynamic Realistic Video World Model with 4D Geometric Control, NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos, SpaceTimePilot: Generative Rendering of Dynamic Scenes Across Space and Time, OmniTransfer: All-in-one Framework for Spatio-temporal Video Transfer): rather than injecting camera geometry through an adapter or a RoPE offset on a frozen video backbone, Atlas rewrites the sequence primitive so camera pose lives inside the AR conditioning stream. The 1-minute-at-1440p operating point is directly comparable to AlayaWorld: Interactive Long-Horizon World Modeling and Google DeepMind’s Genie 3 / Project Genie, and the omni-task framing — 3D reconstruction, video, panoramas, robotics real-to-sim in one model — sits closest to NVIDIA Cosmos 3 (Cosmos 3: Omnimodal World Models for Physical AI) but replaces Cosmos’s dual-tower Mixture-of-Transformers with a single AR-diffusion trunk over a spatial context. The R2S2R sub-claim is the most experimentally testable: if Atlas-driven robot navigation on Marble-format explicit 3D can close the sim-to-real ranking loop that Building Worlds That Train Robots — Real-to-Sim-to-Real (R2S2R) as a Scalable Engine for Training and Evaluating Robot Policies previewed on ALOHA cube-handover, that would be the sharpest existing evidence that a single generative-rollout WFM can serve both the pretraining substrate (World Foundation Models) and the evaluation-layer roles the wiki currently splits across NVIDIA GEAR (SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation), GigaBrain (GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation), and RoboWorld (RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation).