World Labs launched Frameboy, a research-preview demo that brings its autoregressive world model Atlas into real-time interaction: a browser stream from cloud GPUs at up to ~15 fps, so you can fly a virtual camera through a 3D world built from a single image with near-zero retry cost. It is another technical beat after AMD’s late-September ~$8.2B all-stock deal to buy World Labs. Blog: https://www.worldlabs.ai/blog/atlas-realtime. Related: Koala YouRan B++; embodied capital winter.
From Slow Atlas to Game-Boy-Style Frameboy
Atlas, shown ~two months earlier, generates frame-by-frame with pixel-level camera control—image-to-3D worlds or multi-view landmark/site/scout capture—but was slow: moving a camera could take tens of seconds to minutes. After inference optimizations, Frameboy’s retro handheld UI streams new views as you pan/dolly/orbit in the browser. Living-room demos contrast cautious keyframe-wait loops with continuous “drive the camera” exploration. Versus last year’s lighter RTFM preview (mostly interpolating known views), Atlas can invent surrounding scenes from one still or text, and upgrades multi-view novel-view synthesis for browser “cyber tourism” from real landmark photos.
Viewfinder: Build Your Own 3D Playground
Realtime streaming also lands in the web app Viewfinder (closed beta, waitlist): start from your upload, step the camera outward frame by frame; Atlas emits color and depth to keep structure aligned, then one click overlays Frameboy to turn a static build into a live playground. The team casts Atlas as a next-gen generative renderer—while admitting today’s cloud high-end GPUs are still costly vs classic local render. For video-gen realtime/workflow rivalry, see Gemini Omni 1.2 Flash.
$8.2B AMD Tie-Up: Co-Design for Physical AI
On Sept 28 AMD agreed to acquire ~two-year-old World Labs for ~$8.2B all-stock (prior round ~$5B valuation), its second-largest deal after Xilinx; close expected by year-end, with Fei-Fei Li as AMD EVP & chief scientist and Justin Johnson / Ben Mildenhall leading research. Lisa Su wants physical AI—robotics sim, spatial understanding, 3D reconstruction; World Labs needs model–system–compute co-design so realtime generative 3D can eventually leave pricey cloud labs. More: AI News.
FAQ
Q: Is Frameboy a consumer product? A: A research-preview demo over browser + cloud GPUs; Viewfinder realtime remains closed beta.
Q: What’s new vs RTFM? A: RTFM mostly interpolated known views; Atlas can invent explorable scenes from one image/text and strengthens multi-view synthesis.
Q: Why the AMD deal matters here? A: Realtime world models are GPU-hungry; post-deal co-design with AMD silicon/ROCm aims to cut cost and raise speed.
相关文章
