Atlas Unveiled by Fei‑Fei Li: A New Era for World Models and Robotics

Model Overview

Atlas is described as an omni‑model pretrained from scratch to handle four modalities—text, image, video, and 3D—without separate pipelines. It can accept native camera‑pose inputs, enabling pixel‑level control of image and video generation and reconstruction into explicit 3D representations.

Core Capabilities

Camera‑controllable generation : From one to six reference images, the model synthesizes new viewpoints at up to 1440p resolution and produces videos up to one minute long.

Spatial reconstruction : From a single to dozens of input images, Atlas outputs new‑view frames together with point clouds and 3D Gaussian splats.

Spatio‑temporal simulation : By jointly modeling space and time from input video, it supports video reconstruction and Real‑to‑Sim pipelines for robots.

Image generation : Includes text‑to‑image and full‑panorama 360° rendering.

Spatial Context Design

Each input image is anchored in 3‑D space with its camera pose, making camera geometry a native input type rather than a textual description. This eliminates ambiguous prompt‑based camera commands and allows the model to generate smooth transitions between arbitrarily placed reference images.

Architecture

Atlas is a multimodal autoregressive diffusion Transformer. Its backbone combines:

Multimodal : native handling of text, images, camera poses, and depth maps; video is represented as an image sequence.

Autoregressive : outputs are generated element‑by‑element, with each task corresponding to a different sequence permutation.

Diffusion : part of the generation uses rectified flow for step‑wise denoising.

Transformer : ensures compatibility with modern hardware and enables KV‑cache and split‑deployment optimizations inherited from large‑language‑model techniques.

Evaluation

Camera‑controllable generation : Human raters performed blind pairwise comparisons. Atlas achieved win rates of 75 % vs MiniMax H3, 81 % vs Gemini Omni Flash, 86 % vs Alibaba HappyHorse 1.1, 93 % vs FLUX 3, and 94 % vs ByteDance Seedance 2.5. The authors note that competing models lack native camera‑pose input, so the advantage reflects the interface difference as well as visual quality.

Sparse‑view 3D reconstruction : On a benchmark of sparse viewpoints, Atlas attained an average AbsRel error of 25.3 × 10⁻³, outperforming Pi3X (28.7), π³ (34.7), VGGT‑Ω 1B (36.4), Depth Anything 3 (39.3), and MapAnything (47.7). The comparison is limited to the best open‑source dedicated reconstruction models, and Atlas does not lead on every dataset.

Scaling Note

World Labs states that larger compute budgets unlock new capabilities, but model size, training data volume, scaling curves, inference cost, and latency figures have not been disclosed.

Input Quantity vs. Imagination

Blog demos show that with a single ground‑level photo the model generates plausible aerial views filled with imagined elements. Adding a second photo introduces the missing structure while other parts remain fabricated; a third photo aligns the whole scene. Using 2–25 ground photos of Stanford’s main courtyard, Atlas produces a high‑altitude aerial flight path, and with 30 photos it creates an interactive Gaussian‑splat scene.

Robotics Use Case

Two large environments were captured with a phone camera (24 frames each). Atlas reconstructed the scenes and then simulated robots walking different paths. For each step, Atlas generated the RGB and depth observations that a robot’s camera would see, unifying world perception and robot perception in a single model and avoiding error accumulation from separate reconstruction and rendering pipelines.