Hi everyone,
We’re researchers at UC Berkeley working on a tool to make retopology easier and faster. We’ve integrated our tool into a fork of Retopoflow, so it should be pretty easy to learn for those who already use it!
We’re looking for a few alpha testers to try and provide feedback on our tool. Check out a demo here, and sign up to be an alpha tester on our website. Reply below with any questions or comments and we’ll answer!
Thanks,
Rajen and Eshaan
@Metin_Seven @Musashidan @piranha4D @Frozen_Death_Knight I remember hearing each of you talk about your desire for an AI retopologizer, I think. My apologies if this is a mistag 
Sounds exciting! It was just a matter of time before someone started working on this sort of tech.
Is this going to be open sourced, or a closed -source commercial project? How do I sign up, as I couldn’t see an obvious way to do it on the website?
So the AI part is that you draw boundaries that create a patch and it automatically fills that patch in? It’s hard to tell from the demo.
Glad your excited Jamez! The sign up form is the button that says “Join”, with a color palette emoji. We’ll update the website to make it more clear 
.
This project will probably be a mix of open source and closed source. We want to open source as much of it as we can, but there will be parts that are hard to open source like a dynamically growing database. We plan to charge some amount of money for it in the future, as the add-on relies on a pretty compute-intensive backend!
Yeah, the algorithm that autofills patches is the AI part! Our initial implementation is inspired by a SIGGRAPH paper from a few years ago, and we are improving on it with more data/better algorithms.
Thanks for the info. You can probably imagine I had already tried the join link but you get a page that says “You need permission
This form can only be viewed by users in the owner’s organization”
Oops! Thanks for letting me know–I’ve fixed it now.
It’s great to see someone using AI to create tools like this, instead of pretending to create art. 
Retopo and unwrap are two of the areas I would have thought would be a primary target for these machine learning tools.
I envisioned a tool that is a mix of what you’re doing here and something like Wrap 3, but based on the primary facial muscles/structure FACS system with the goal of creating perfect deformable geo.
Marking out these major loops with curves, and the computer filling in the patches. This could work for any creature and not just a humanoid, once these primary patches are blocked in.
hehe you forgot me there 
I’ve been wanting this for ages, but I’m not entirely convinced this is the right path.
Something where you draw a few guides (as Danny shows) and it tries to create the loops and close the gaps between would be useful.
This seems a lot like the patch tool that was already in Retopoflow, what is the source it was trained on?
Also, will you train on people’s use of it in the beta, and: will it send potentially NDA’d assets to you?
Lastly, will beta-testers get a license afterwards?
Is it this paper?
There’s a lot of amazing research that we unfortunately never end up seeing in Blender, it’s nice to see someone breaking this trend.
Yes, I believe Topogun 3 and maybe even 3DCoat have similar tools like this to fill in the patches. A system like this ideally needs to be trained around facial muscles/expressions, and the FAC approach.
Thanks for your comment! I completely agree, and what you described is pretty much what we were envisioning for the future of the tool as well. Although it is quite impossible to do now, our goal is to collect enough data to enable powerful features like this in the future.
We’re currently using data extracted from 40 models from the Blender open source movies. To the best of our knowledge, there are no large open source datasets of 3D models with good topology, which is why nobody has solved retopology yet. A big goal of ours is to collect data through the tool so that we can create better algorithms and get closer to fully automated retopology. We haven’t thought much about NDA’d assets, but if it is a problem we can definitely avoid storing the asset. Alpha testers will be super useful to us, and will absolutely get a license!
Yes! And I think Topogun 3 and 3DCoat use an earlier paper by the same authors called Pattern Based Quadrangulation.
Core RetopoFlow dev here, and I am assoc prof of CS at local university. Interesting work! Mind if I join in on the conversation?
Hey Jon, we’d love to hear your thoughts!
Make sure you abide by the license of the open movie assets, which I believe is Creative Commons Attribution. And make sure you’re explicit about using alpha/beta testers data to train your model. AI training is right now famously bad at attribution and licensing, and it’s easy to do it better!
I believe there’s a class action lawsuit going on in America at the moment, against whatever the names of the two main art stealing programs are.