World Labs launches Atlas for camera-path world generation

• Atlas makes camera trajectory a native input rather than a vague cinematic instruction. The launch claims up to one minute of 1440p video, a specification venturetwins' Atlas announcement summarized with the model's text, image, and camera-path inputs.
• A few ordinary phones can become a post-production capture rig: venturetwins' reframing demonstration shows three fixed phone angles reconstructed into a moveable new viewpoint.
• Sparse references buy speed at the cost of inference. bilawalsidhu's post on sparse reconstruction points to the appeal of capturing only what needs to be exact, while Atlas generates the rest of the scene.
• Atlas can turn its reconstruction into explicit point clouds or 3D Gaussian splats, not only rendered video, according to World Labs' launch post.

World Labs' own demo includes a one-minute path through a scene built from a small reference set. The same post says two unrelated images can be positioned in a shared 3D context and joined with invented connecting spaces. Radiance Fields' technical rundown places the output in the Gaussian-splat pipeline already used by World Labs' Marble product.

Camera geometry is Atlas's central control surface. World Labs says each image or depth map is attached to an explicit camera pose, so a creator can specify position, angle, and a manually designed route through the generated scene instead of relying on terms such as pan or crane in text.

The company describes a camera-controlled workflow in three parts:

1. Supply one to six reference images, or build a scene from text.
2. Place the references in the model's shared spatial context.
3. Design the camera path, then generate shots along it.

Atlas can produce new views beyond the frame of a reference image. In the camera-control examples, World Labs uses a single robot image to generate the robot's unseen back side and a surrounding lawn. The creative shift is control at the shot-design layer, while the off-camera world remains a model prediction.

World Labs' clearest filmmaking example is a bullet-time-style reframe. Its researchers recorded a subject with three to five ordinary phones or action cameras on tripods and backpack-sized clamps, reconstructed the scene, then moved the virtual camera through it.

The before-and-after clip starts from three fixed views and ends with a continuous zoom and pan through the reconstructed space. bilawalsidhu's four-iPhone footage called out the unusual part: an orbiting free-viewpoint result from four iPhones, where comparable volumetric capture once called for far denser camera coverage.

World Labs says the same space-time mode can freeze an event for reframing. That makes a captured take editable as coverage, not merely as a locked camera angle.

Atlas does not distinguish between a reconstructed fact and an unseen region it had to infer. Its sparse-reconstruction section is unusually direct about that trade-off: one photo can yield a new aerial view, but portions outside that photo are imagined.

Adding input images narrows the ambiguity. World Labs says two or three images can often make a reconstruction faithful, while the model can accept more than 100 images when a location needs closer recreation. In its cottage example, a garden remains consistent after one photo, a second photo adds the cottage, and a third fills in the main house.

A Hacker News discussion similarly separated Atlas's high-quality novel views from a physical simulator or robot planner. Atlas can supply scenes and sensor-like outputs for a separate robotics workflow, but the launch does not claim that it directly predicts robot actions.

Rendered frames are only one end of the pipeline. Atlas natively handles 2D image frames and 3D depth maps, then exports a reconstruction as a point cloud or 3D Gaussian splat, which World Labs says can render on-device at high resolution and frame rates.

That output matters for workflows that need a scene asset rather than a video: World Labs names gaming, design, VFX, and robotics. As Radiance Fields explains, Atlas differs from World Labs' earlier RTFM model by adding explicit 3D output, although its splats still include filled-in regions that no camera captured.

Atlas is not a public editor or API at launch. World Labs says it is entering early access with select partners and is collecting access requests, while future versions of Marble and other World Labs products will use the model.

World Labs has not published pricing, latency figures, or named partners in the announcement. The launch post includes no paper, model card, or code release either, as Radiance Fields noted.