We show more results on out-of-distribution input videos, such as model generated videos, and videos with extremely large object or camera motion.
[Back to main page]

Given a monocular video, naively fitting a 3DGS on it leads to broken results due to the low coverage of the scene. Instead, we first run 360Anything to synthesize the entire 360° view as a panorama video, and then train 3DGS on it. This leads to much better geometry and enables fly-through exploration of the 3D scene.
[Back to main page]

We show the input perspective videos, the text prompts, and the panorama videos generated by Imagine360 [1], Argus [2], ViewPoint [3], and our 360Anything. Red boxes highlight the area corresponding to the input video.

• Imagine360 and Argus produce videos with low visual quality and clear artifacts.
• ViewPoint always places the input frame at the center of the output. It generates significantly rotated panoramas when inputs contain large camera motions, leading to severely distorted objects.
• Our 360Anything generates high-quality canonical 360° panoramas and follows the text prompt well.

[Back to main page]

360Anything also applies to perspective-to-panorama image generation. We fine-tune FLUX on synthetic 360° rendering of 3D scenes, which generalizes to in-the-wild and OOD images.
[Back to main page]