[c-hendricks]

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.

[valine]

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

[Jolter]

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

[valine]

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

[TillE]

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.

[greysphere]

"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.