I’ve been making content for over a decade and a half and have never really had eyes on my personal work. On retrospect I have had over 500,000 ears my music but also didn’t know until the audience around it had already moved on, you can read more about that here if you’re interested. So instead of working on my own content I’ve spent the last couple of years creating for some of the biggest tech and gaming influencers/ brands with already established audiences.
After watching an episode of Corridor Crew where they’d turned their friends into illustrated characters, I just got a feeling that I could figure out how they’d done it, which lead to the creation of this video. And although I’d worked on content for clients that had reached the millions in terms of viewership, it was still amazing to see something that I’d made off my own back had netted over 200K views in a couple of days and had taken me from 120 subscribers to over 4000.
The tech evolves so quickly that I’m told the first video is a little dated now. This one focuses around the use of Google Collab (if you don’t have a decent PC), but I quickly realised my 3090TI could handle Automatic1111 so I quickly moved over to a new workflow for training which I showcased in Part 2. If you are struggling with the changes on the Collab doc since this tutorial, I recommend following my ‘Part 3’
Part 2 does a recap on the intial setup, how to take images and also information about resizing. It also covers the installation of Automatic1111 and then this time, instead of Dreambooth, I show you how to utilise Hypernetwork. It’s another form of ML training. Although I’d argue it isn’t as good as Dreambooth. Still definitely one to check out if you’re looking to experiement with alternativer methods of image training.
In my Part 3, I showcase another Google Colab method of Dreambooth training by TheLastBen, I’ve been told people have had a lot more luck getting this method to work compared to the methods in Part 1. I also share a little bit of an inpainting trick to paste faces onto the genrated characters if you’re struggling to get an image spot on straight away.
These tools really are incredible, it’s a little bit full circle for me going back to alternative reality faces. My first job in the gaming industry was motion tracking at a video game animation studio that eventually got bought by Epic Games. Some of my manual laboring is likely still part of the dataset that is now used for the Metahuman creator. It’s definiely something I intend to explore further and as these 3 videos head towards half a million total views, I suspect it’s something a lot of people are interested in too.
If there’s anything you’d like me to explore in terms of A.I faces or animation, be sure to leave a comment in one of the YouTube videos!
