World Labs' Atlas: camera-controlled video that reconstructs in 3D

Fei-Fei Li's startup opens early access to a model that exports 3D you can use. Plus: Gemini 3.8 Flash reveal today, Perplexity runs local on Mac.

Nowline SEP 2 6:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • It generates worlds, not clips

    Atlas is a multimodal world model, not a text-to-video toy. Feed it native camera poses and it renders up to a minute of 1440p video you can pan, truck, and crane through, plus 360° panoramas — the camera is a real input, not a phrase you hope lands.

  • 3D you can actually export

    From one image to dozens, Atlas reconstructs a scene into point clouds and 3D Gaussian splats and does novel-view synthesis from sparse inputs; push 100+ images in for a faithful capture. That's exportable geometry you can drop into a 3D pipeline, not just frames.

  • What you could build this weekend

    Bullet-time reframing of a clip, camera-controlled establishing shots, or Real-to-Sim for robots: Atlas emits RGB plus depth sensor data for navigation and models rigid, articulated, and deformable interactions. A weekend game or sim prototype now has a spatial backend.

  • It won the head-to-heads — but it's gated

    Testers preferred Atlas 93% of the time over FLUX and 86% over Gemini Omni Flash on camera-controlled generation, with 3D error of 25.3 vs 28.7 for the nearest specialist. The catch: it's early access with select partners via a Typeform — no public pricing or GA date yet.

  • Elsewhere: Google reveals Gemini 3.8 Flash today

    Google DeepMind is set to publicly reveal Gemini 3.8 Flash (codename 'skimaki') today, reportedly a refinement pass aimed at less-verbose coding output at 3.7's pricing ($0.75/$3.75 per M tokens). Whether API access lands with the reveal is unconfirmed — watch the endpoint.

  • Elsewhere: Perplexity keeps sensitive prompts on your Mac

    Perplexity's new Hybrid Compute auto-classifies prompts and routes sensitive ones to a local model on Apple Silicon, sending the rest to the cloud — no token fees for the local runs. If you work with confidential data, that's a privacy gate you didn't have to build yourself.