Hacker News new | past | comments | ask | show | jobs | submit login
3D Hair Synthesis Using Volumetric Variational Autoencoders (linjieluo.com)
95 points by juliendorra 5 months ago | hide | past | web | favorite | 16 comments

On the topic of hair synthesis... a particularly mind-blowing legal complaint alleges a startup CEO that has published and shown at Siggraph repeatedly of fabricating large elements of their tech. It was brought by a core research scientist working on hair (who brought the complaint) who also alleges that he was fired without warning when he revealed this fabrication (after consistent warnings that he would do so if changes were not made at the company), and that the CEO and others assaulted him while attempting to keep him from removing personal data from his laptop. Other gems: the research scientist alleges that they withheld months of pay, that the CEO doesn't even have a work visa valid for his work as CEO, and that the CEO used his position as a professor to copy results from papers under review and integrate them into their product. If they don't settle this will be a very interesting lawsuit.


And take a look at that, the founder in question is the last author on this paper. And another member of Pinscreen is the first. I'm going to have fun replicating this one.

I remember that Real Time Live presentation. It was... bad. Watching it again, I just noticed this gem of a comment from when it was first posted to YouTube:

"That avatar maker around 37:00 seems too good to be true." ~vfxforge, 1 year ago


This company Pinscreen fakes results in academic publications which their former VP of engineering documents: http://sadeghi.com/dr-iman-sadeghi-v-pinscreen-inc-et-al/

Well that was a hell of a read. I'd go as far as to say that anybody who's checking into this thread should read it.

It just kept escalating.

My main observation is that the main players in this have very strange hair for people whose field of research is hair rendering...

That is just fascinating. Thanks for the link.

Oh god. Some of those "input images" are very clear signs that the whole thing is faked.

It's funny how the volume model for the dog predicts an empty space for the shoulders, which the dog doesn't have. A hiccup of creating it from a trained model of hairstyles, I suppose.

Static is nice but next step would be inclusion of CG physics to realistically simulate hair motion while head is being rotated quickly (meaning lots of fun affine transform minutiae incorporating inertia). There's probably a market for that in anime.

Is it just me or do the volumetric renditions not even come close to matching the source photos? It's like the photos were picked later to try and match the graphics.

Oh that dog is amazing. I wish they had more of those examples.

Almost makes me want to implement the paper, just so i can add long flowing locks to my dogs :D

The results look blurry to me, particularly the top left blonde hair. I think Pixar has achieved better looking results.

Interesting to see Snap's involvement here. I wonder what kind of Snapchat integration they're considering.

Oooh! I was wondering how you do a volumetric autoencoder.

So is it fake/half fake/genuine?

Applications are open for YC Summer 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact