1) It was mostly manual, though AIs were useful in certain filtering tasks.
2) Depends on the quality you are seeking. If you only want one run of a similar, off-the-shelf model, around the 1000s is enough. But at the number of iterations you have to run to build your own and improve results, you probably need about 100k.
To tackle this problem, we built our own supercomputer off of parts we bought off of ebay, though I can't say I recommend that route, because it now lives in our living room.
Haha, this is something that constantly ponder, while I am developing this product.
Though, the conclusion I've come to is that that hand-drawn art will always meaningful for humans, because it is born of the human experience.
An interesting example is the invention of photography, which at its time, was very good at doing the thing artists were doing back then (capturing likenesses)
But photography didn't replace art: instead, artists now use photographs to be more expressive, convincing, and make better art. In tandem, the widespread adoption of photography meant that more average folks could get their likenesses taken!
Personally, my skills as an artist has improved by quite a bit, after launching this product, purely because observing it offers some fascinating insights into how anime is created!
I hope that as an industry, we'll find better ways to create, and what we know to be the "best" art today will be even better in the future!
It's an extremely hard research problem, because darker skin tones account for only about 0.3% of all anime art produced in the world.
We have employed an absolutely exhaustive array of art and data science tricks to give the model the ability to draw darker skin tones, though they are underrepresented. The results that you see today are the culmination of many months of careful tuning!
It's not definitely perfect, but from a data science perspective, this situation can't be rectified until the art world makes a shift.
Personally, I hope that more art representing dark skin tones will be created in the world!
also: If the discriminator starts off perfect, then the generator can't learn to be better.
Sort of like a human learning to play chess: If you start off with top-tier opponents that crush you, then you don't have a gradient to learn from. Instead, you need players at your own level to grow your skills.
Ah yes, the fine line between charming anime character and lovecraftian horror
There was such popular demand for these "horror" images that we made them part of the generation in V2! If you refresh enough on the webpage, you can find some horrors!
We're currently working on the data migration from V1! As long as you are using the same email as you did in 2019, you'll be able to see the image again!
As for a V2 generation, sorry, because the models are different, you'll have to discover a similar image again, if you want a V2 version!
You can think of it like coordinates on a many-dimensional vector grid.
We craft the functions the functions that will illuminate sets of those points based on a combination of observation, what we know about our model architecture, and how our data is arranged.
And yes, when the model is retrained, we have to discover them again!
I hope you all enjoy playing with the new and improved generator! We've been hard at work improving the model quality since the last time the site was posted[1]
As both a professional fantasy illustrator & software engineer, I find the concept of AI creativity so fascinating. On one hand, I know that mathematically AI only can hallucinate images that fit within the distribution of things that it's seen. But from the artist perspective, the model's ability to blend two existing styles into something so distinctly new is so incredible (and not to mention also commercially useful!)
Anyways, happy to answer any question, thoughts, or concerns!
https://www.gwern.net/Danbooru2020
Though now we have made our own :)