4 ms·
> If people figured out how to generate this kind of art without exploiting uncompensated unwilling artists' free labor, it would be a different story. No it w
by visarga 4y ago
> If people figured out how to generate this kind of art without exploiting uncompensated unwilling artists' free labor, it would be a different story.
No it wouldn't. It would still compete against artists. We'd have worse models in the beginning and it would take time until someone licensed enough images to improve the models, but the capability is there and we know about it, too late to stop.
By the way, Stable Diffusion has been fine-tuned with Midjourney image text pairs. So now we also have AI trained on AI images.
- orbital-decay 4y ago>So now we also have AI trained on AI images It doesn't matter, and never did in the first place. All large models (including SD) are already trained on other models output, since there's simply no possible way to have a high quality tagged dataset of the size they need. Smaller models are used to classify the data for larger ones, then the process is repeated for even larger models, with whatever manual data you have. Humans only select the data sources, and otherwise curate the entire bootstrapping process. This kind of curated training actually produces better results.