``Splatter Image'' is now available that allows you to create images and animations that look like they are viewed from 360 degrees around you based on a single image

Splatter Image, a renderer that outputs the 3D shape of an object in an image from just one input image, has been released.

Splatter Image

GitHub - szymanowiczs/splatter-image: Official implementation of `Splatter Image: Ultra-Fast Single-View 3D Reconstruction'
https://github.com/szymanowiczs/splatter-image

'Splatter Image' is based on 'Gaussian splatting,' a technology that performs 3D composition in real time based on images taken from multiple viewpoints, and is the first to apply Gaussian splatting to 3D composition from a single image. It is explained that it was done.

You can see an example of the generated animation below.


Normally, in Gaussian splatting, points and voxels are projected onto a 2D image, associated with a Gaussian function, and the information of the points and voxels is expressed on the 2D image while being expanded. Splatter Image predicts hidden elements, such as wheels or chair legs, that are hidden in a single image and assigns points to the appropriate locations to represent the hidden parts.

Prediction requires model training, but when trained on car objects, it was completed in 10 minutes on a single GPU and showed excellent results in standard benchmarks.


This method is touted for its cost advantages over alternative methods.