> ## Content Index > Fetch the complete content index at: https://www.implicator.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Fei-Fei Li's World Labs Launches Atlas but Withholds Paper, Price, and Partners - URL: https://www.implicator.ai/world-labs-atlas-withholds-paper-price-partners/ - Published: 2026-09-02T03:37:24.000Z - Updated: 2026-09-02T03:37:24.000Z - Description: World Labs launched Atlas, a world model that generates up to a minute of camera-controlled 1440p video from a few photos, into early access with unnamed partners. No paper, price or model card came with it, and every benchmark win was measured by World Labs itself. - Author: Marcus Schuler - Tags: AI News World Labs [launched Atlas](https://www.worldlabs.ai/blog/atlas?ref=implicator.ai) on Sept. 1 into early access with unnamed partners. It can generate up to one minute of camera-controlled video at 1440p from a small number of photographs. The launch arrived with no paper, no model card, no price, no named partner, and benchmark results that only World Labs has run. What Changed - World Labs launched Atlas on Sept. 1 into early access with unnamed partners, with no paper, model card, price or general-availability date. - Atlas generates up to one minute of camera-controlled 1440p video from a small number of photos and outputs point clouds and 3D Gaussian splats. - Every benchmark result is World Labs' own: rivals in the camera test got text prompts while Atlas got native camera paths, and the company reran all reconstruction baselines itself. - One baseline, VGGT-Ω 1B, carries an Aug. 18 notice from its authors saying its results may be inflated by benchmark contamination; the Atlas post does not mention it. AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/). ## What Atlas does Atlas combines text, images, camera poses and depth maps in a shared spatial context, meaning each photo is pinned to a position in 3D space and the model generates whatever comes next from that arrangement. Its camera-generation examples use one to six input images with manually designed camera paths. A separate one-minute 1440p example used a “small number of reference images.” It can reconstruct scenes from as few as two or three images, then output point clouds or [3D Gaussian splats](https://radiancefields.com/what-is-gaussian-splatting?ref=implicator.ai), a lightweight 3D rendering format for ordinary hardware. The launch also shows Atlas reframing footage recorded with three to five phone cameras. Atlas will power future versions of Marble, but the new model is available only through an early-access application. ## World Labs set the test In World Labs’ camera test, 75% of third-party raters chose Atlas over MiniMax H3 and 94% chose Atlas over Seedance 2.5\. Atlas received native camera trajectories. Its rivals received text descriptions of the intended movements. The launch post concedes that “more sophisticated prompt engineering or creative multimodal prompts could improve camera following for some models.” FREE WEEKDAY MORNING BRIEFING Don’t miss the next AI story that matters. The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern. Email address Send me tomorrow’s briefing Check your inbox for the confirmation link. About five minutes. No hype. No spam. Across seven reconstruction datasets in the same company test, Atlas recorded average absolute-relative point-map error of 25.3, compared with 28.7 for posed Pi3X, about 12% lower. On Tanks and Temples, Atlas tied Pi3X at 42.4 and trailed VGGT-Ω at 40.2. No independent replication of either benchmark has been published, and World Labs has not said whether it will publish the evaluation code, dataset splits or model outputs. ## A baseline warning The Meta AI and Oxford VGG repository for VGGT-Ω carries an [Aug. 18 notice](https://github.com/facebookresearch/vggt-omega?ref=implicator.ai) saying benchmark contamination in an ancestor checkpoint means the released 1B model’s results “may be inflated.” It asks evaluators, “please do not rely on them until we conclude our investigation.” World Labs is reporting a win over a baseline whose own authors have said its published results may be inflated and should not be relied on while their investigation continues. Atlas launched 14 days after that notice. Its post does not mention the warning or identify which VGGT-Ω checkpoint World Labs reproduced. ## The missing product details No paper, arXiv entry, model card or code accompanied the launch. World Labs disclosed no parameter count or training-compute figure and described the training material only as “a large diverse corpus of multimodal data.” Know someone who'd find this useful? [✉️ Email it to a friend in one click](mailto:?subject=A%20newsletter%20I%20think%20you%27d%20like&body=This%20is%20one%20of%20maybe%20three%20newsletters%20I%20actually%20read.%20The%20rest%20just%20pile%20up%2C%20unread%2C%20judging%20me.%0A%0AAnd%20yes%2C%20this%20email%20mostly%20wrote%20itself%2C%20which%20is%20a%20little%20on%20the%20nose%20for%20an%20AI%20newsletter.%20Doesn%27t%20make%20it%20wrong.%20implicator.ai%20is%20good.%0A%0ASubscribe%20free%3A%20https%3A%2F%2Fwww.implicator.ai%2Fsubscribe%2F%3Futm%5Fsource%3Dnewsletter%26utm%5Fmedium%3Dforward%26utm%5Fcampaign%3Demail%5Fforward), or they can [subscribe free here](https://www.implicator.ai/subscribe/?utm%5Fsource=newsletter&utm%5Fmedium=forward&utm%5Fcampaign=forward%5Fto%5Fcolleague). As of Sept. 1 evening, the World Labs [API documentation](https://docs.worldlabs.ai/api/models?ref=implicator.ai) listed four Marble models and no Atlas entry. Atlas has no disclosed price, general-availability date or named early-access partner. ## The simulation boundary Atlas “aids in building” robot simulations, the launch post says. Its two robotics environments were reconstructed from 24 phone-video frames each in the demos. Physics, policy training and evaluation belong to a separate [Real-to-Sim-to-Real engine](https://www.worldlabs.ai/blog/real-to-sim-to-real?ref=implicator.ai) that came with World Labs’ July 21 acquisition of SceniX, founded by Yunzhu Li and Changxi Zheng. World Labs has not specified which parts of those demonstrations run on Atlas. Native camera poses are a substantive addition, and explicit 3D output goes beyond RTFM, the learned renderer World Labs announced in October 2025 without an explicit geometric representation. Atlas also comes from a team whose co-founder Ben Mildenhall co-created NeRF, a method for building novel views from images. Vincent Sitzmann, who leads the Scene Representation Group at MIT CSAIL, was speaking about whether video-generation models trained on pixels are the route to physical intelligence, the approach World Labs and its rivals share, when he told [Ars Technica](https://arstechnica.com/ai/2026/07/simulating-everything-sort-of-the-promise-and-limits-of-world-models/?ref=implicator.ai) in July: “I’m saying it’s a bet. This is not a decided thing.” Frequently Asked Questions What is Atlas? Atlas is World Labs' new world model. It combines text, images, camera poses and depth maps in a shared spatial context and can generate camera-controlled video, reconstruct scenes from a few photos, and output point clouds or 3D Gaussian splats. It is in early access with unnamed partners and will power future versions of Marble. Did Atlas beat rival video models? In World Labs' own camera-control test, 75% of third-party raters chose Atlas over MiniMax H3 and 94% chose it over Seedance 2.5\. Atlas received native camera trajectories while rivals received text descriptions, and the launch post concedes better prompting could improve the other models. No independent replication has been published. What is the VGGT-Ω contamination warning? VGGT-Ω 1B, one of the reconstruction baselines Atlas is compared against, carries an Aug. 18 notice in its Meta AI and Oxford VGG repository saying benchmark contamination in an ancestor checkpoint means its reported results may be inflated. The Atlas post, published 14 days later, does not mention the notice. What has World Labs not disclosed about Atlas? No paper, arXiv entry, model card, code, parameter count or training-compute figure accompanied the launch. Atlas has no disclosed price, general-availability date or named early-access partner, and the World Labs API documentation listed only Marble models as of Sept. 1. Does Atlas simulate robots? The launch post says Atlas 'aids in building' robot simulations by reconstructing environments and generating sensor views. Physics, policy training and evaluation belong to a separate Real-to-Sim-to-Real engine that came with the July 21 SceniX acquisition, and World Labs has not specified which parts of the demos run on Atlas. AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/). [Alibaba Turns Happy Oyster Into Real-Time AI World Model for GamesAlibaba Group released Happy Oyster on Thursday, a new AI world model for generating interactive 3D environments, Bloomberg reported. The model can create video worlds that users steer while they are The Implicator![](https://www.implicator.ai/content/images/2026/04/2026-04-16-04.18.49-happy_oyster@2x.webp)](https://www.implicator.ai/alibaba-turns-happy-oyster-into-real-time-ai-world-model-for-games/) [Yann LeCun Raises $1 Billion for AMI Labs in Europe's Largest Seed RoundLeCun's AMI Labs raised $1.03B at $3.5B valuation for world models. Europe's largest seed targets robotics, healthcare, manufacturing.The Implicator![](https://www.implicator.ai/content/images/2026/03/2026-03-10-05.57.05-robot_billion@2x.webp)](https://www.implicator.ai/yann-lecun-raises-1-billion-for-ami-labs-in-europes-largest-seed-round/) [World Labs Raises $1 Billion From Nvidia, AMD, and Autodesk for Spatial AIWorld Labs, Fei-Fei Li's spatial AI startup, raised $1 billion in a round backed by Nvidia, AMD, Autodesk, and Andreessen Horowitz, the company said on Wednesday. Autodesk put in $200 million, the sinThe Implicator![](https://www.implicator.ai/content/images/2026/02/2026-02-18-12.37.12-world_labs_spatial@2x.webp)](https://www.implicator.ai/world-labs-raises-1-billion-from-nvidia-amd-and-autodesk-for-spatial-ai-2/)