Thoughts on AI Image Generation (Part I)
I enjoyed Nina G. 's thought-provoking/ wide-ranging discussion on AI/ML(artificial intelligence/machine learning)?image generation. Thought I’d add my thoughts on the topic as a separate article.
Nina posited that the “5 most common arguments against AI-generated art” are?(my responses immediately below)
?I’ve not heard this argument from a visual artist or a photographer. I think any painter trained in a technique like egg tempera or a photographer familiar with the daguerreotype process understands that creating art using an AI image generator requires similar rigor, subtlety and sophistication in approach.
Nina described her own process as including images generated from dozens of text prompts, multiple AI software tools, plus work designed in Figma and post-processed Photoshop.in?It’s the digital equivalent of mixed media & it’s impressive.
In art school I had a professor who told us?“Michelangelo would have used an airbrush.” I get the argument and agree that artists incorporate the technology of their day.
A text prompt for an image generator certainly reduces the “barrier-to-entry” if you will for creating an image, but it’s not like you just type in “Mona Lisa” and get a Mona Lisa, so to speak. (Though you might get THE Mona Lisa)?
I?focus my post on dry subjects such as Bias, Copyright, Commerce, Datasets,?and the Wheel - the last two will be covered in a subsequent article.
A Very Brief Intro to AI/ML tools
Very, very broadly speaking, AI/ML tools - for both image or text processing - go through a three step process of modeling, dataset training and inferencing in order to do the voodoo that they do. So an image generator has been trained on bazillions of images, including thousands of images of bananas and thousands of images of castles, so that when you type in “Banana Castle” into the text prompt for an image generator you get an image result somewhere between a bright yellow Windsor Castle and a Cavendish with battlements.
Bias
Any algorithm-based technology runs the risk of producing output that contains inherent bias. This is not a new issue - Google search perhaps being the most publicized example from years past, with documented real-world impacts (see article here, based on research here) Journalists and technologists are starting to see a similar trend with AI image generators as well (see article here, referencing in part work done by Dr. Sasha Luccioni with Stable Diffusion on HuggingFace here)
In short, this means when you type in “CEO” into the prompt of an AI image generator, you’re likely to get images of a white, middle-aged male.
Copyright
Copyright is an incredibly complicated subject, but for the sake of brevity let’s group it into two sections:?copyrighting AI-generated art (which I’ll cover here) and copyrighted art used to train a model - I'll cover that in Part II
Important: there is no such thing as global copyright laws - they vary from country to country.
In the US, the Copyright Office ruled last year that you can’t copyright AI generated art (see article about it here, with the original second request for reconsideration to the Copyright Review Board here)?
So what this could potentially mean is that anyone could download Nina’s cool AI-generated image, make a poster of it to sell on Etsy for $199.99. Her work, being AI-generated, is in essence copyright-free, so anyone in the US can use it, even for commercial purposes and not have to compensate her. Nina, perhaps a little aggrieved, would have hire a lawyer and prove that there was “substantial human input” on her part (all those Figma files and Photoshop tint corrections) that gives her a right to copyright the work and go after the Etsy-posting poster bandit that monetized it.
?Last year AI Image generator OpenAI?offered full usage rights of the output of it tools to users to reprint, sell, and merchandise (see article here), which is great, but not much of an offer if those images are copyright free already. I guess we can just hope that in the future an AI image generator company doesn’t claim that they own the output as well. (Although StarryAI is already going down the watermark route)?
Using AI image generators for commercial purposes (Such as Cosmopolitan’s recent magazine cover) is a new trend, but it’s one that’s keeping a lot of technologists and lawyers up a night (or at least the journalists at TechCrunch) Nina mentioned the use of AI generators to (re?)create signatures - I don’t even want to go there.
In short, the copyright issues for AI-generated art are unclear, very complicated and potentially very litigious. This is a subject that will make a few law firms a lot of money in the near future.
Commerce
In our highly-developed, mixed-market economy, one cannot separate art from commerce. Takes money to make art, and art is often made to make money. When talking about AI-generated images, we need to think about how the AI tech companies are making money, how companies can use AI images to save money & is there room for an artist to use AI to make commercial or fine art.
How the AI companies make money?
AI companies are as we speak trying to figure out their pricing model &?how to grow their business. Many are following the Software-as-a-Service (SaaS)?playbook, refined over the years by companies like Adobe, AWS, Salesforce, Vistaprint etc. The first step is typically offering the software for free to gain an initial set of users, then introduce some sort of pricing structure, which can span from individual purchases via credit card to enterprise accounts with seven-figure 2-year subscription pricing.
OpenAI’s going with a usage model as is Stability.ai. Midjourney has a subscription model, which I bring up because it mentions GPUs. Running AI software requires a lot of expensive processing power such available via GPU chips.
<!— Shameless plug—>
Smart AI companies?of course look to my company Neural Magic to get GPU-class performance on less-expensive CPUs :)
<!— /Shameless plug—>
In short: the era of free-image generation via browser is closing; if you want to use AI to make art, be prepared to pay $ for a platform subscription, for image output, or both.
Can companies can use AI images to save money?
Yes. A creative director at a magazine company can hire an human illustrator to do a cover for (average going price) $2,000 or buy a $30/mo subscription to an AI generator and have their unpaid art school intern do the cover in the afternoon.?
You do the math.
The creative fields - especially illustration - have been massively transformed (battered?) by digital tech disruption since the 1990’s. Fifty years ago, there was a thriving illustration industry across print, publishing, advertising and signage. Now the field barely exists. People draw picture yes, and there are a few superstar artists, but’s it similar to the music industry: they’s Taylor Swift, and then there’s everybody else - one person at the front of the long tail making tons of money, and a lot of people making very little to nothing off their art.?
Smart graphic designers felt the way the wind was blowing in the 2000’s and transformed themselves into UX designers. But even they shouldn’t get complacent. With out of the box design systems, SaaS tools and icon libraries abounding, they are a cost center ripe for reduction. As I heard a chief creative officer tell his 50-person design staff at an unnamed company, a couple years back, “You will all be replaced by AI.” (True story, really)?
Will AI image generation lead to a new breed of “imageers” that use AI, code and design to create new images and new worlds? Yes, definitely. Do we know how big or how lucrative such a career path might be? We don't yet know.?
Can commercial artists use AI to make commercial art?
Yes, though they probably should be aware of just inputting a text prompt doesn’t qualify as “making” and take steps to ensure they are compensated and their image rights will be protected. ?
(End of Part I; To be continued...)
Business Broker
2 年This is a good read, Rob. If AI can render a banana yellow Windsor Castle, then yes it is taking everyone's jobs ??