Comment by euix | Hacker News Reader

Comment by euix | Hacker News Reader

Midjourney has been really helpful to me as a one man dev. I can mockup art much faster then what I can do in photoshop. I still intend to at some point do a complete pass using a professional artist (or learn to draw myself) - because generative art is not consistent thematically from asset to asset. But if I just want to see what my tile assets look like if they were all done in 30's art deco style, I can do it in 20 minutes.

As placeholders or to create little bits and doodles (like a mouse cursor in the style of an armored fist), there are lots of little graphical icons in a game that would other have to be created by a graphical artist. Generative art is really useful in my experience.

It's reduced the work to the point where I can toy with it in my off time and spend most of my effort in the actual programming and development.

The other idea I have toyed with, coming from professional ML experience - was to build my own generative model and use it to create my own art assets. Here I wonder how the copyright rules would work - would the assets I train on be subjected to copyright? This is a much bigger conversation at that point and I wont be the only one affected.

And no one is trying to take that away from you. You are using it as a tool, and intend to pay someone to create the final version, or do it yourself.

The issue people have is when you just use a dataset trained on someone else's work and pass it off as your own, and in the case of Steam games, most likely profit from it.

What if I look at other people's art and learn from it? Seem unfair to pass that work off as my own. All artists should be banned from looking at copyrighted images, we can't risk them incorporating copyrighted elements into their own work. /s

That’s not a good analogy and you should know it.

It’s a terrific analogy. The alternative is to believe that a 5GB model somehow contains a database of 160 million images.

"With four parameters I can fit an elephant, and with five I can make him wiggle his trunk”

It seems like you could calculate how much data is within X% error of a 5GB model, and what X% should be for 'visual data'.

I bet it's pretty big.

It's a fine analogy, but the map is not the territory. Machine learning is not human learning, even if it works in a vaguely comparable way.

It's still a computer program that uses an enormous amount of copyrighted work as its input.

Yeah I don't see too much issue in using generative art for more trivial things, like some banner art to sit atop a blogpost or something. Placeholders also seem like a really good application, particularly for cases where the randomization might expose issues that real users would face. I wouldn't have spent money on these things anyway.

Direct incorporation of generative art into a commercial product is much more murky.

"The other idea I have toyed with, coming from professional ML experience - was to build my own generative model and use it to create my own art assets. Here I wonder how the copyright rules would work - would the assets I train on be subjected to copyright?"

Unless you're training on assets available in the public domain, any generated output from your "custom model" would have the same potential copyright issues as midjourney, stable diffusion, etc. What exactly are you confused about?

I think programmers also often confuse "public domain" with "publicly accessible on the internet"

Reply on news.ycombinator.com