4 ms·
Lots of people in here making arguments about the fact that the way these image models learn is roughly analogous to how people learn, the fact that these relat
by humanizersequel 4y ago
Lots of people in here making arguments about the fact that the way these image models learn is roughly analogous to how people learn, the fact that these relatively tiny models simply don't have enough bits to grok anything except the most popular (and therefore recurrent in the training data) images, etc.
What about the fact that these models aren't just randomly spitting out and taking credit for random images? This seems the most salient point to me — if I used a paintbrush to create a copyright-violating clone of some notable artwork or IP and tried to pass it off as my own, I'd be breaking the law. We wouldn't try to ban paint and canvas and the human arm because it has the potential to create something that infringes on copyright, we'd enforce the actual act.
If these models make this kind of infringement easy, then they are bad products and their users will run the risk of going to court. The whole thing seems like a non-issue.
- shlubbert 4y agoStable Diffusion 2.0 has already made changes to make infringement harder by removing names of artists and celebrities from the dataset, which is probably the right move -- now if you want to emulate e.g. Greg Rutkowski you actually have to figure out how to describe his style in the prompt, teaching you a little about what makes his art special and making it easier to create your own style along the way.
- CuriouslyC 4y agoOr get a bunch of Greg Rutkowski paintings and perform textual inversion to get an embedding of his style, which is what people are actually doing, and go back to square 1.
- humanizersequel 4y agoIf the result of this is an image that infringes on his copyright — I'd imagine this would have to be a pretty 1-1 copy, since his style appears to very generic — then the burden should still be on the person who produced the image and published it.
- MrNeon 4y agoCelebrities and artists were not removed from the data used to train SD 2.0. 2.0 uses a new text encoder trained from scratch and it just did not capture the same famous names as the OpenAI CLIP used in 1.x.