Comment by sillysaurusx | Hacker News Reader

Comment by sillysaurusx | Hacker News Reader

Parent Full thread

Cartoony generation is probably a matter of training on the appropriate source material, for what it’s worth.

https://dreambooth.github.io shows a glimpse of the future. You’ll be able to upload a few drawings that you want to emulate (e.g. Mickey Mouse), and then you can give it a specific prompt (e.g. Mickey Mouse doing a handstand).

You’re probably right about the consistency of 3D concepts, though. On the other hand, I was going to say “If you need a specific table, AI might not be able to help” — but again, dreambooth shows that we might be able to upload a few photos of a certain table, and it’ll take care of the details.

Give it a few years. :)

I think you’re spot on that AI will be an incredible tool for artisans. I used it to make some video game music: https://soundcloud.com/theshawwn/sets/ai-generated-videogame... Even though I can’t play any instruments too well, I was able to craft each piece uniquely. (My favorite is “Crossing the Channel”, which has a strange rhythm because I’m pretty sure the AI made a mistake at the beginning, and then extrapolated the next “actually, this isn’t a mistake” song that it thought of, which turned out to sound cool. A bit like a guitarist doing improv.)

I think the main issue with cartoony generation is that professional animators, are not being tasked with "here is famous character Mickey Mouse that's omnipresent in your source data, now draw him on the moon wearing a hat", they're being tasked with something where the source material is a couple of concept images possibly at lower fidelity than the desired end output (and the task might well be much more specific, and the art directors considerably more pedantic about the quality and consistency of the lines than the average hacker typing in magic incarnations to get something resembling fantasy art for their blog). Of course, if you want to make memes of Mickey Mouse wearing a hat on the moon, or Mickey Mouse at the White House with a hammer and sickle, or an anime-style Mickey Mouse, AI cartoons may well be plenty good enough already.

There's a role for AI in filling in gaps, but illustrators remain essential in creating the basic style to extrapolate from and even more so in professional quality work (And to some extent, other procedural generation techniques were able to fill gaps up before cutting-edge NNs - the stampede of wildebeest in the 1994 Lion King was procedurally generated from a handful of models for example)

DreamBooth works with just 4 input images.

Agree that dreambooth is the future. I'm building an app that lets users with no technical background train their own dreambooth models for $2-$4: https://synapticpaint.com/dreambooth/info/ They can also share their trained models for others to use.

Here are some potential use cases:

- for fun (giving yourself a makeover, inserting yourself into famous movies)

- cheaper way to get studio photos (wedding photos, professional headshots for actors/models)

- easy way to create marketing assets. Like if you own an etsy store and don't want to engage a marketing studio, instead just create a dreambooth model of your necklace or whatever and create high quality product photos

There are probably a bunch of other use cases. I think making this easy (no figuring out how to do a git pull or rent a gpu) plus the community sharing aspects will make this technology a lot more accessible to artists and general users, and then the users will be doing all kinds of cool things with it organically.

Reply on news.ycombinator.com