The Colorado State Fair hit global news last month. It wasn’t because of some weird animal or corndog incident. It was down the very concept of what we call art. That’s because an Artificial intelligence (A.I.) created artwork walked away with first prize in the Digital Arts segment of the Fine Arts competition. this has led to a lot of outrage and navel gazing in the digital Arts world.
And turns out A.I can do photos now too. But should we worry too ?
If you’ve click on this post you’ve already seen one A.I created photo like image. The street does not exist nor does the night market below. These were generated on webpage on my mobile phone in about 60 seconds from a simple one line prompt.
If you look closely you can see some errors in both images but it take you time to notice and at first glance you can easily be fooled into thinking they are kosher.
But they were both created by Artificial Intelligence using a simple text prompt that I entered on my mobile’s web browser
So it’s not just Art the A.I. can do photos
So What Exactly happened in Colorado ?
As Vice reported on the 29th of August, Jason Allen, the owner of a tabletop games company walked away with the first prize in the Digital Arts Category at the State Fair.
No great shakes I hear you say. Not my cuppa tea but he has talent. But to some the problem is how he created it – he used an A.I image generator called Midjourney
And things went a bit crazy.
“Art is dead, dude. It’s over. A.I. won. Humans lost.”
Jason hasn’t exactly helped by saying quotes like ths to the NY times. But there has been lots of posting about the doom of art or that it was a cheat to enter.
To be fair he did create it using a tool and his prompt. It’s likely he will have selected the image over a few botched one and may well have run it again through Midjourney. He then polished that image in photoshop and blew the image up with Gigapixel before printing it onto Canvas. so not exactly a passive partner.
There was some criticism that he didn’t let the Judges know enough. He did use the “Jason Allen via Midjourney” on his pieces but never explicitly said that involved A.I.. However as the NY times reported the judges would still have awarded him the prize.
But there is fear and loathing. Some of that may be driven by the potential impact of this on digital artist. Particularly those that rely on small commissions to meet their needs. And there is a point. Say I want an illustration for my new book cover, I can now DIY it with much less skill and much less time. And no artist to pay
But I betcha physical media artist were saying the same thing when Digital artistry took off. And photographers were met with contempt and fear back in the 1800’s for much the same.
Art will live on but it has been a cold new dawn for some.
I’m sorry Old Chap let me put down my paper and tell me what is this A.I. Trickery ?
We’ve become use to A.I. help with digital editing. Be Honest, many of you will use your photo editing software to adjust level, contrast etc with varying degrees of A.I. support and remove dust etc even off your digitalized analogue photos.
And the big buzz over the last few years has been over A.I. supportered face swap technology, from dodgy Porn deepfakes, to image manipulation in the news to playing around swapping yer mate’s face onto Boris.
And the tech behind this is all pretty clever but you give the computer 2 images and get it to map over a small portion.
The rise of A.I. Image generation
But now the A.I. doesn’t need images to work from a simple prompt will do. The woman below doesn’t exist
A.I. image creation is done by machine learning. There are various systems. But in the end these all boil down to the systems reviewing millions of image images and linking those to text prompts. How the AI construction the image can be varied using a random number sequence (generated by a so called seed. If you keep the seed the same you’ll get a identical (or near depending on the software) image.
This Dall-E mini image set illustrates this as all generated for me from just one prompt “Photorealistic Red Apple”
And what’s more these tools can run in a web browser on a mobile phone. No need for a powerful PC .
Power to the People or Uncanny Valley ?
And the Ai is getting slicker.
The image below was created in just one run of a person that does not exist (outside Mario games).
It can get things wrong at time badly but items are improving. And it’s worth noting these images were run off in a web browser. You can do this on a cheap PC, Chromebook or even your smartphone. In fairness these public accessible A.I. are not without issues.
Sometimes you get a great image first go. The big challenge has been photo realistic hands and eyes as this image demonstrates
But the same A.I. chucked this out later.
And if you think that still has an uncanny valley feel take a look at the photo below.
Ghosts in the Machine ?
The woman in the 2019 photo does not exist. The image was created by StyleGAN a very advanced A.I. system developed by NVidia Researchers. Although this system is not full proof it produces much more realistic images. For more example visit thispersondoesnotexist.com (hit refresh to change the image and watch out for the odd oops )
Dall-E 2 is also being billed as the A.I. image generator to put this into the mainstream.
And also these advanced facial A.I. are now available commercially. companies like Icon8 produce and sell these online. They get used for advertising (no model or photographer fees) and a host of legitimate uses. They even allow you the ability to make your face from a simple features menu (here’s one I did but they charge you to download it)
Okay So What are the issues ?
(1) They suffer from (our) biases at bit
There is an issue within current A.I. and that is bias. In part that’s down to us. The A.I. use massive amounts of images we create to form their algorithms and sadly we tend to have marked stereotyping.
I ran just the prompt “Nurse” in Stable diffusion and got this
Now the A.I did a good job here face wise. But they all look like white females. I ran it 3 times again. It was only on the third batch of images that I clearly get a nurse of black and minority ethnic (BAME) background. There were 2 other image in the whole run whom may have not been caucasian but it was hard to tell. And guess what all 36 images appear to be of female nurses.
That’s pretty awful. NHS England identified that 1 in 5 nurse are of a BAME background. And whilst there are issues about gender inclusion in nursing over 1 in 10 nurse in the UK are male. But not according to the 36 image in a row I made via stable diffusion. Not one male nurse.
…and the cause
And that bias is down to us. We’ve made more stereotypical images over the years
Interestingly google has put in algorithms to produce much more diverse image on search So when I searched for nurse on Google images I got a much more diverse set of images in both race and gender.
And it’s not that A.I. algorithms can’t do darker skin tones and yes the nurse generated was spontaneous, but really it shouldn’t require you to race prompt.
Interestingly Google are not allowing public access to Imagen, their A.I generator in part due to this (although oddly they’ve fixed it on their image search). They noted “..the data requirements of text-to-image models have led researchers to rely heavily on large, mostly uncurated, web-scraped datasets. While this approach has enabled rapid algorithmic advances in recent years, datasets of this nature often reflect social stereotypes, oppressive viewpoints, and derogatory, or otherwise harmful, associations to marginalized identity groups.”
(2) Potential for Misinformation and fraud
And yes sadly these are tool that can be used for misinformation. and yes these engines are incredibly powerful allowing state agents, agitators and a whose of folk to to potential use them for mischief or worse.
A few prompts and you can have a fake moon landing. Or worse
I’ve not spent ages doing that but you get the idea how good the software can be. It does good likeness of well known people and AI can be trained.
It’s not just image misinformation. AI created images are being used to create false accounts. In 2009 facebook took down over 900 accounts with AI generated images.
But this isn’t anything new. A.I assisted face swapping technology existed long before Art generators. And before that well there was always photoshop favoured tools of fraudsters agent provocateurs and meme creators Alike.
But Photoshopping goes back to the dawn of photography. The Stalin’s USSR was well know for making folk disappear from official photographs. But it dates back further. A good example is this photograph US Civil War General and later President Ulysses S. Grant. It’s his head but it’s not his body nor was the rider in the original image
(3) No point the A.I. can do it
Well yes A.I can produce stunning images and the errors will get better, but photorealistic images predate the digital era and never killed of photography.
For decades digital artist have produced photorealistic images but in fact photorealism goes back to the 1960’s. And photorealism & digital art has bet genuine photos in winning contracts but so does photography pinch some of their markets. A.I. is never going to catch the intimacy of wedding photos, family holiday shoots or, that perfect day in the studio/on the mountains etc.
Yes it will provide another choice to companies for their images but it’s not the first time artificially created images have been used.
But could A.I. help photographers?
There is a good chance you use A.I to tweak your images in the the likes of lightroom, photoshop and Luminar. The new AI image generators now can tweak images. You can mask areas such as clothes and get the generator to change those.
Most spectacular is something called outpainting. You can take a core image but then expand the edges of this in frames gradually building up a much bigger image as shown here
The tech is likely to incorporated into improving editing. You can also use the image to create the basis of another
A final thought on A.I. photos –
More Opportunities than things to fear
It’s a new media and it brings great challenges and issues. But it also brings opportunity and I suspect we have much less to fear than we could benefit from it. But it may be tougher in the Digital Arts world, although again may be a gamechanger for others.
And funnily enough folk were saying the same thing when D
Wanna Play with A.I. ?
I think you’re better trying these, than being worried. There are several programs available but i’ll mention the key ones. I’m focusing on the big generators that you can access and this is not an exhaustive list.
All the tools that are available can be run in a web browser.
(1) Dall-E 2
Arguably the biggest name at the moment, Dall-E 2 offer potentially game changing features like realistic eyes, allowing you to outpaint an image. And I’d love to play with it but I sit on a waiting list with millions over at open.ai
Their Dall-E 2 page has an interactive limited demo and the chance to join that list too. Dall-E 2 generates 4 images per credit. New users currently get 50 free credits and every month get 15 more free. More credits can be bought. The headline figure you read of $15 for 115 credits (460 images) but some report you are encourage to buy way more.
(2) Craiyon (formerly Dall-E mini)
Craiyon is the most accessible and fun of the tools. Open source and completely free to use. But it is the weakest of the tools here. It produces 9 images in a go. But it really struggles with faces and images have a lo res quality. But it’s free and fun.
(3) Stable Diffusion
Stability.ai’s Stable diffusion is current major rival to Dall-E 2 and is useable without a wait. Funnily enough, most images here were created by this A.I. . What’s more there are several platforms using it.
Most Notable are Stability.ai own version which I guess is bleed edge stuff with the most up to date stuff. You get currently 200 free credits. Everytime you generate a 512×512 image on basic settings you use a credit. Make the image bigger, make multiple, use more processing time the cost goes up.
For casual users who want to dabble with no commitment it offers the cheapest block purchase of credit options (1000 cost just a tenner). You can set up a subscription as well which may be cheaper if you use in the long run.
NightCafé also gives you access to Image diffusion, but also 3 other A.I. art orientated image generators. You can generate an image in one and use in the others to evolve that image. It gives 20 free credits for new accounts but can get 5 free each day if you log in and claim. And you can get more free credits by engaging in its social content . You can pay but it’s more pricey ($14.99USD for 100 credits). It also offers subscription. It has a much better interface plus it stores your images for you.
I include Midjourney for completeness. It after all is the A.I. that caused all the fuss in Colorado. It does more art orientated images than photo. Again it offers a free trial of 25 credits. But unlike the Stable diffusion offerings you must sign up to subscription for more credits rather than by a block
(5) Google Imagen
Potentially on of the most powerful A.I. image generators on the planet. But you won’t be using it properly any time soon. As Engadget reported – Google worries about how A.I. generators can be exploited to make inappropriate or malicious images and the potential for bias.