WonderHuman: 3D avatars from single-view video

This link (which is not underlined and so easy to miss, boooo) includes videos and more: https://wyiguanw.github.io/WonderHuman/

Not sure if there are similar techniques applied, but this definitely reminds of of the AI "microwave" filter that's been very popular on TikTok for the past few weeks. It creates equal numbers of impressive and hilarious 360 models from single images. One example: https://www.tiktok.com/@beccasthesia/video/74714552131157558...

Reminds me of that scene [1] from Enemy of the State where they "Rotate us 75 degrees around the vertical" in one of those "enhance, enhance" tropes.

[1] https://youtu.be/3EwZQddc3kY?t=7

Based on the video this looks great in the 200x200px previews typical for a paper but has substantial visual artifacts that become obvious at higher resolution.

Maybe there are straightforward ways to improve the results by spending more compute time. But as presented it's more "turn one picture into a convincing character in your Sims game", not "turn one picture into an animated character in your next TikTok video"

Looking forward to the next paper on this. Hopefully debuting the FaceBack app!