Atlas - A world model for spatial intelligence

Hi everyone! Give Atlas one image and it has to imagine. Give it a few more, and the world starts getting constrained. It’s interesting because every image is treated as a view from a specific point in 3D space. Atlas takes those views together and tries to infer a world that could make all of them true at once. Anything outside them is still up for interpretation. Add another image and that world may change again. You can even put two unrelated images into the same spatial context and let Atlas figure out what could exist between them. Maybe a hallway, a doorway, a corner... whatever makes both views belong to one continuous place. From there you can move the camera through it, generate up to a minute of 1440p video, or reconstruct the scene as explicit 3D. We only ever see one slice of a world at a time too, and somehow keep inferring what probably exists beyond it. So... I guess anything is possible? :)