Few examples from the demo. You can generate various styles from a single prompt. You can either pick semantic matching images, or not and get unexpected results.

[1] unsplash.com/photos/a-boa...
[2] unsplash.com/photos/mount...
[3] unsplash.com/photos/man-i...
[4] unsplash.com/photos/macro...

This can be used for multiple applications such as color matching or diffusion guidance. Here, we showcase the diffusion process of generating a medieval house with the reference to the right.

Variance in MC is quite common in computer graphics so we combined ReSTIR -- more precisely the weighted reservoir sampling -- with SWD to keep more impactful random directions in the optimization.

Happy to announce: ReSWD. Sliced Wasserstein Distances are quite powerful, but they perform a Monte Carlo (MC) integration (over random directions). During an optimization this can lead to noisy gradients due to variance.

Project page: reservoirswd.github.io

One neat implication is that we can edit the point cloud to fix missing features or wrong scaling. We even created a small gradio component for simple edits in the demo (pypi.org/project/grad...)