Hook
Most video models dream up pretty frames and forget them the moment the camera moves. Atlas, launched September 1 by Fei-Fei Li’s World Labs, does the opposite: it generates the video and commits to real 3D geometry for everything it just showed you. One pass, one model — you get a filmed camera path and the Gaussian splat scene underneath it.
The Story
World Labs has been the spatial-intelligence company to watch all year. Their first product, Marble, turned a photo into an explorable 3D world. Atlas is the generation that comes next — and it is not a Marble update. It was pretrained from scratch as a single multimodal model that speaks text, images, video, and 3D natively.
Under the hood it is a multimodal autoregressive diffusion transformer — a rectified-flow model that works on both 2D image frames and 3D depth maps at the same time. In plain terms: it thinks in pictures and in space together, instead of guessing the space after the fact. Feed it as few as two or three photos and it reconstructs a faithful scene; hand it a hundred and it holds them all in context.
The trick that matters for us is camera control. Atlas takes camera geometry as a native input. You draw the path, it generates plausible new frames along that exact path — up to one minute of 1440p video — and then bakes the result into explicit 3D. In their Stanford Main Quad demo, 2 to 25 ground-level snapshots became a smooth aerial fly-over the photographer never actually shot.
The output is not a black box you can only re-render through the model. It reconstructs scenes as point clouds and converts them into full Gaussian splat scenes — the same format that now imports into Unreal, Unity, Blender, Rhino, and half the tools on your drive. That is the whole point: generate once, then own the geometry.
Why You Should Care
Two things separate Atlas from the flood of “world model” demos this year.
- It closes the loop between video and 3D. Video models give you frames. Splat capture gives you geometry. Atlas gives you both from the same forward pass — no photogrammetry stage, no separate reconstruction tool.
- Camera paths are an input, not a happy accident. For anyone who storyboards shots — VFX, arch-viz, virtual production — controllable camera geometry is the feature that turns a toy into a tool.
- It reconstructs from a handful of images. Two or three photos to a faithful scene means real, messy, under-shot locations become usable 3D.
- The output is editable splats. You are not locked into World Labs’ renderer. Drop the scene into your DCC and keep working.
The honest caveat: this is an announcement, not a repo. At launch there was no paper, no arXiv, no model card, and no code — Atlas is entering early access with select partners. So treat the numbers as World Labs’ own, and the demos as demos. But the direction is real, and it lines up with where the whole field is heading: generation and reconstruction becoming the same act.
Try It / Follow Them
- Read the announcement: worldlabs.ai/blog/atlas
- Request early access: via the World Labs site — they are onboarding partners now.
- Follow along: @theworldlabs and Fei-Fei Li on X.
- Context and coverage: Radiance Fields’ breakdown.
IK3D Lab Take
We have covered World Labs three times because they keep shipping the thing everyone else is still promising. Atlas is the clearest statement yet of their bet: the future is not a video model or a splat capture tool — it is one model that does both, and hands you geometry you can actually edit. Early access with no code out yet means we are watching, not deploying. But put a controllable camera and editable splats in the same forward pass, and you have described the pipeline a lot of us have been faking with three tools and a prayer. The day the API opens, this is the first one we try.
The key distinction is that Atlas doesn’t just generate a camera move—it outputs the explicit 3D splat behind it. That commitment to persistent geometry could make “revisiting” a generated scene far more practical than models that only produce pretty frames.
The key distinction is that Atlas doesn’t just generate a camera move—it outputs the explicit 3D geometry behind it. That “3D splat” could make the result far more useful for scene editing and consistent viewpoint changes than models that only produce attractive frames.
The key innovation is that Atlas doesn’t just generate a convincing camera move—it outputs the underlying 3D splat as well. That commitment to explicit geometry could make the result far more useful for editing, navigation, and downstream 3D workflows than a video-only model.
The key leap is that Atlas doesn’t just render a convincing camera move—it outputs the explicit 3D splat behind it, giving creators something persistent to inspect and reuse instead of disposable video frames. That could make camera-controlled world generation far more practical for production workflows.
The camera move plus the explicit 3D splat is the compelling part here—Atlas isn’t just generating plausible frames, it’s committing to geometry that can persist as the viewpoint changes. That could make “video generation” far more useful for actual 3D workflows.
The explicit 3D splat is the compelling part—Atlas doesn’t just render a convincing camera move, it preserves the geometry that makes the move possible. That could make World Labs’ approach far more useful for editing and interactive 3D work than models that only output pretty frames.
The key leap is that Atlas doesn’t just generate a camera move—it outputs the explicit 3D geometry behind it. That commitment to a reusable 3D splat could make the result far more useful than a video model that only produces pretty frames.
The key leap is that Atlas doesn’t just render a convincing camera move—it preserves the explicit 3D geometry behind it as a usable splat. That could make the transition from generated video to editable 3D worlds far more practical.