>100 Views
November 20, 26
スライド概要
本スライドでは、World Labsが公開した空間知能(Spatial Intelligence)のためのオムニ世界モデル「Atlas」の技術詳細と関連研究を解説しています。
Atlasは、テキスト・画像・動画・3Dデータを単一の空間コンテキスト上でゼロから事前学習したマルチモーダル・自己回帰拡散トランスフォーマー(Multimodal Autoregressive Diffusion Transformer)です。これまで別々のモデルで扱われていた「カメラ軌道制御付きの動画生成(最大1分・1440p)」「1〜100枚以上の写真からの3D再構成(点群・3D Gaussian Splats)」「時間を止めて視点を動かすバレットタイム等の4D時空間シミュレーション・ロボット用RGB-Dデータ生成」「360度パノラマを含む画像生成」の4領域を1つのモデルで統合しています。また、3D再構成ベンチマークではVGGT-Ω 1Bを上回る精度(mean pointmap AbsRel 8.6 vs 9.7)を記録しています。
In this presentation, we explore "Atlas," an omni world model for spatial intelligence introduced by World Labs.
Pretrained from scratch on text, images, video, and 3D within a single shared spatial context, Atlas is built upon a Multimodal Autoregressive Diffusion Transformer architecture where each image and depth map is paired with its camera pose. It unifies four distinct capabilities into a single model: (1) camera-controlled image and video generation (up to 1 minute at 1440p), (2) spatial 3D reconstruction from 1 to 100+ photos into point clouds and 3D Gaussian splats (outperforming VGGT-Ω 1B with a mean pointmap AbsRel of 8.6 vs. 9.7), (3) 4D space-time simulation including bullet-time reframing and RGB-D data synthesis for robotics, and (4) 2D image and 360° panorama generation.
東京大学発のベンチャーです。AI人材育成およびDeep Learningを中心とした先端AI技術を活用したソリューションを提供するAI総合研究所として活動しています。
Tech Discussion Atlas: An Omni World Model for Spatial Intelligence World Labs · Technical blog · September 2026 NABLAS Research/Engineering Community
Atlas: one world model for generation, reconstruction and simulation Introduction World Labs' omni world model, pretrained from scratch on text, images, video and 3D in one shared spatial context. Camera-Controlled Generation Spatial Reconstruction Space-Time Simulation Images or video along Real scenes from 1 to Bullet-time reframing Images and 360° a designed camera 100+ photos, as point of video, and RGB-D panoramas from text path, up to 1 min at clouds and 3D simulation for robots or image prompts 1440p Gaussian splats Image Generation World Labs, Atlas blog (Sep 1, 2026) © NABLAS Inc. All Rights Reserved 2
Camera-controlled generation follows a designed camera path Application 1/4 • Input: 1‒6 reference images plus a hand-designed camera path • One image: the whole scene from new angles, e.g. the back of a robot • Spatial context: two unrelated images placed in 3D, with the transition generated • Long video: 1 minute at 1440p ▶ Live demo in the blog World Labs, Atlas blog · Camera-Controlled Generation © NABLAS Inc. All Rights Reserved 3
Application 2/4 Spatial reconstruction scales from one photo to a hundred • Input: 1 to dozens of photos, no special capture rig; 100+ also fit • Fidelity: 2‒3 images usually suffice; more images, less imagination • Output: point clouds and 3D Gaussian splats, same format as Marble • Benchmark: mean pointmap AbsRel 8.6 vs 9.7 for VGGT-Ω 1B (World Labs' eval)in the blog ▶ Live demo World Labs, Atlas blog · Spatial Reconstruction © NABLAS Inc. All Rights Reserved 4
Space-time simulation: bullet time and robot data Application 3/4 • Reframing: 3‒5 phone or action-camera views; • Robot navigation: scene from a casual phone video freeze time and move the camera • Manipulation: vary objects, positions, motion, (24 frames); generates the robot's RGB and depth ▶ Live demo in the blog lighting and background World Labs, Atlas blog · Space-Time Simulation © NABLAS Inc. All Rights Reserved 5
Application 4/4 Image generation is a secondary, but capable, skill • Input: text or image prompts • Output: images and 360° panoramas • Strengths: complex prompt following, text rendering, many visual styles • Focus: the blog calls this secondary to spatial tasks ▶ Live demo in the blog World Labs, Atlas blog · Image Generation © NABLAS Inc. All Rights Reserved 6
What the blog discloses: a multimodal autoregressive diffusion transformer Text, images, camera poses and depth share one Multimodal Elements are generated one at a time from earlier ones. Each task is a different input → output sequence. Diffusion Transformer Engineering context. Each image and depth map carries its camera pose; video is an image sequence. Autoregressive Technical details From LLM serving KV caching, cache-aware routing, disaggregated serving From latent diffusion Distillation, CFG, shifted noise schedules, better VAEs Rectified flow. Fewer denoising steps trade quality for speed. Matmul-heavy and hardware-friendly; quality keeps rising with compute. Not disclosed Parameters, data, compute, speed World Labs, Atlas blog · Technical Details © NABLAS Inc. All Rights Reserved 7
CameraCtrl: camera poses as per-pixel Plücker rays Related work 1/3 • Idea: turn each camera pose into a per-pixel ray map (o×d, d), like a 6-channel image • How: a camera encoder injects it into the temporal attention of a frozen video diffusion model • Also used by: LingBot-World 2.0, a 14B causal world model (Plücker via AdaLN) CameraCtrl framework (He et al., ICLR 2025, Fig. 2) • Relation to Atlas: Atlas ties every image and depth map to a pose; ray maps are the common way to do it He et al., CameraCtrl, ICLR 2025 · Robbyant, LingBot-World-Infinity, 2026 © NABLAS Inc. All Rights Reserved 8
VGGT: feed-forward 3D reconstruction in one pass Related work 2/3 • Idea: a plain transformer maps 1 to 100s of images to cameras, depth, point maps and tracks • How: DINO tokens + camera tokens, alternating frame-wise and global attention, < 1 s • Follow-ups: VGGT-Ω scales to 10B and dynamic scenes; VGGT4D, PAGE-4D, V-DPM extend it to 4D VGGT architecture (Wang et al., CVPR 2025) • Relation to Atlas: Atlas also outputs depth and pose, and beats VGGT-Ω 1B on pointmap error (8.6 vs 9.7) Wang et al., VGGT, CVPR 2025 · VGGT-Ω, CVPR 2026 · VGGT4D, PAGE-4D, V-DPM (2025‒26) © NABLAS Inc. All Rights Reserved 9
Autoregressive diffusion: one sequence, any split of given and generated Related work 3/3 • Transfusion: one transformer, next-token loss on text and diffusion loss on images sequence order → Generate Text I₁P₁ P₂ Î₂ P₃ Î₃ • Diffusion Forcing: each token has its own noise level, so any subset can be context or Reconstruct I₁ I₂ I₃ P̂ D̂₁ P̂ D̂₂ target P̂ D̂₃ • RTFM: World Labs' earlier AR diffusion Simulate I(v₁,t) given I(v₂,t) I(v₃,t) P*, t* Î transformer over posed frames D̂ generated (I image, D depth, P pose) • Relation to Atlas: the same model can generate (few views) or reconstruct (many views), as in the diagram Transfusion, ICLR 2025 · Diffusion Forcing, NeurIPS 2024 · World Labs, RTFM, 2025 · diagram: presenter's illustration © NABLAS Inc. All Rights Reserved 10
Summary Summary 1 2 3 One model covers camera-controlled video, 3D reconstruction, 4D simulation and image generation The blog discloses a multimodal autoregressive diffusion transformer in which every image carries its camera pose Each capability has its own line of work (CameraCtrl, VGGT); AR diffusion may be what lets one model combine them World Labs, Atlas blog · related work as cited © NABLAS Inc. All Rights Reserved 11