Thanks for an entertaining yet important deep dive! I enjoyed reading both chapters and am looking forward to part 3.
Though I'm afraid that AI models are more capable of mimicking specific artists and their styles than it appears at first glance.
I think the Midjourney team might have intentionally "nerfed" this capability in V4. This is also mentioned in the "Midjourney: 6 months later" video you've linked to, around the 8:20 mark, and one can argue this is good news for artists who don't want to see their names used wholesale.
At the same time, models like Stable Diffusion do a better job at it to this day. Here's how SD did with your "Norman Rockwell" prompt: https://dreamlike.art/d/A5mlKgBPiABc (seems way closer than what Midjourney spit out).
Already back in September last year, I did an experiment to see how 5 famous artists would deal with unconventional subject matter they wouldn't have actually worked on:
To my---granted, extremely untrained and amateur---eyes, Stable Diffusion does a pretty decent job of mimicking each artist's style quite convincingly.
Of course, I took some of the most well-known and the most "trained on" artists from this list: https://www.urania.ai/top-sd-artists. So you might be right that not every artist has something to fear in the immediate future.
But I wouldn't underestimate the potential of upcoming algorithms to better mimic any style convincingly with even less training data.
I think we should inevitably expect more regulation in this space.
I personally would like to see "opt in" / "opt out" options for artists who want to be excluded from training data, as well as some form of a profit-sharing model that allows participating artists to get royalties if AI images / videos / etc. based on their style are monetized.
Things are moving fast. Let's see how it all develops!
Thanks for an entertaining yet important deep dive! I enjoyed reading both chapters and am looking forward to part 3.
Though I'm afraid that AI models are more capable of mimicking specific artists and their styles than it appears at first glance.
I think the Midjourney team might have intentionally "nerfed" this capability in V4. This is also mentioned in the "Midjourney: 6 months later" video you've linked to, around the 8:20 mark, and one can argue this is good news for artists who don't want to see their names used wholesale.
At the same time, models like Stable Diffusion do a better job at it to this day. Here's how SD did with your "Norman Rockwell" prompt: https://dreamlike.art/d/A5mlKgBPiABc (seems way closer than what Midjourney spit out).
Already back in September last year, I did an experiment to see how 5 famous artists would deal with unconventional subject matter they wouldn't have actually worked on:
https://www.whytryai.com/p/ai-art-experiment-5-artists-3-prompts
To my---granted, extremely untrained and amateur---eyes, Stable Diffusion does a pretty decent job of mimicking each artist's style quite convincingly.
Of course, I took some of the most well-known and the most "trained on" artists from this list: https://www.urania.ai/top-sd-artists. So you might be right that not every artist has something to fear in the immediate future.
But I wouldn't underestimate the potential of upcoming algorithms to better mimic any style convincingly with even less training data.
I think we should inevitably expect more regulation in this space.
I personally would like to see "opt in" / "opt out" options for artists who want to be excluded from training data, as well as some form of a profit-sharing model that allows participating artists to get royalties if AI images / videos / etc. based on their style are monetized.
Things are moving fast. Let's see how it all develops!
thanks for writing this!