DeviantArt, the “largest online social network for artists and art enthusiasts,” announced DreamUp, an AI-powered text-to-image generator service powered by Stable Diffusion. The new service has seriously angered some members of the community, so much so that DeviantArt announced that all hosted images will be “automatically labeled as NOT authorized for use in AI datasets.”
This is a temporary victory for the luddites. I can assure you – speaking as one of the first Local 802 members (the musicians union) to be listed as an “electronic music synthesist” – this is not a future the angry artists can count on. That said, this nonsensical “John Henry” argument raises an interesting question: Where will training data for AI come from?
There are two general answers:
- Data collected from real-world things, events, and behaviors.
- Synthetic data (aka artificially-generated information).
The DeviantArt community feels their artwork was/is being used to train a Stable Diffusion-powered text-to-image generator without their consent, yet when each and every one of the angry artists trained to become artists, they scraped every piece of art from every possible source, using every ounce of associated metadata to become the artists they became. They did this (in most cases) without the consent of the original artist. This human training technique is as old as humans; we (human beings) are the greatest mimics on earth.
You could argue that the humans in training were doing something different than DreamUp is doing. They weren’t. Learning is learning whether it’s a human or a machine.
What about the business model? That’s different, isn’t it? No, it’s going to do what it can to commercialize what it knows.
Stable Diffusion learns just like we do, only it’s not as smart. (“Smart” is not the right word here, but using jargon would not help communicate my meaning.) Perhaps it would help if you saw a few million of the images and associated metadata from the LAION (Large-scale Artificial Intelligence Open Network) dataset used to train Stable Diffusion. As you can see from the training data, the artwork must have associated metadata (data that describes the artwork) to be useful. As you can easily imagine, images with better metadata are more useful for training. (Note: If you have installed your own Stable Diffusion model, here’s a really nice, very simple tutorial for fine-tuning it with your own images.)
Let’s imagine a few probable futures.
- Angry artists have stopped the training of AI models. We can cross this off the list… it’s never happening (but angry artists can dream).
- Angry artists can contribute their art to training sets with consent (for compensation). It’s easy to imagine a business there, as well as an additional AI partnership business to help create volumes of synthetic data.
- The most probable future: the best artists will never think about big training sets; they’ll craft their own fine-tuning training sets (as noted above) and start to create insane, as-yet-unimagined artwork based on their own use of this new and magical tool.
The lesson: don’t get angry… get creative!
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it.