New Text-to-Art Image Generator Easier to Misuse, Critics SayRelative to DALL-E 2, it is easier to use to generate pornography, as controversial website 4Chan has recently demonstrated
Stable Diffusion, a new open source text-to-art generator (they’re coming thick and fast), dropped a couple of days ago:
It has come under fire for some of its more questionable uses. Most uses have been reasonable, according to Kyle Wiggers at TechCrunch, but the problem uses give us a window into potential headaches to come:
On the infamous discussion board 4chan, where the model leaked early, several threads are dedicated to AI-generated art of nude celebrities and other forms of generated pornography.
Emad Mostaque, the CEO of Stability AI, called it “unfortunate” that the model leaked on 4chan and stressed that the company was working with “leading ethicists and technologies” on safety and other mechanisms around responsible release. One of these mechanisms is an adjustable AI tool, Safety Classifier, included in the overall Stable Diffusion software package that attempts to detect and block offensive or undesirable images.Kyle Wiggers, “Deepfakes for all: Uncensored AI art model prompts ethics questions” at TechCrunch (August 24, 2022)
But, you guessed it, Safety Classifier can be disabled and Stable Diffusion, which can work with art of public figures, has no filters for pornography. The 4chan website has generated controversy for most of its two-decade existence so tumbling to this use is perhaps not a surprise. .
Is it illegal? Maybe, but…
If history is any indication, however, enforcement will likely be uneven — in part because few laws specifically protect against deepfaking as it relates to pornography. And even if the threat of legal action pulls some sites dedicated to objectionable AI-generated content under, there’s nothing to prevent new ones from popping up.Kyle Wiggers, “Deepfakes for all: Uncensored AI art model prompts ethics questions” at TechCrunch (August 24, 2022)
Given that DALL-E 2 and Midjourney are both grabbing attention just now, this controversy may at least make the public more aware of both burgeoning text-to-art software and potential deepfakes.
Stable Diffusion claims to be better than DALL-E 2 (claims examined here and here.)
The differences between the three, compared, are quite revealing. Analyst Fabian Stelzer notes,
“Midjourney reminds me of a beautiful analog Moog synthesizer — it’s almost impossible to make it sound bad and you can do incredible things with it, but in exchange, its range is more limited. The artifacts it does are like analog distortion, very pleasing.” …
Stelzer says that perhaps the best-known AI image generator, DALL-E 2, is like “a digital workstation synth — incredible range, but it almost always sounds a bit too digital.”
“Stable Diffusion is like a complex modular synthesizer, you can get almost any tone out of it, but it’s a bit harder to play and prompt.”Matt Growcoot, “AI Image Generators Compared Side-By-Side Reveals Stark Differences” at PetaPixel (August 22, 2022)
Look at the text-to-images of the same subject that he chose for comparison and see what you think:
DALL-E 2 vs Midjourney vs StableDiffusion mega thread: photography, illustration, painters, abstract
these image synths are like instruments – it’s amazing we’ll get so many of them, each with a unique “sound”
rules: same prompt, 1:1 aspect ratio, no living artists pic.twitter.com/47syy7uPJJ
— fabians.eth (@fabianstelzer) August 20, 2022
Note: The featured image is AI-generated art.
You may also wish to read:
The threat that deepfakes pose to science journals Image manipulation has been a problem for decades but convincing deepfakes could magnify the problem considerably. When Xiamen University researchers tested their deepfakes on experts, two of three fared poorly by eyesight alone; one scored 60%. Not reassuring.
New OpenAI art program does NOT claim copyright for AI. As DALL-E 2, which generates blended images in response to key words, moves into the art world, a key question has just been settled. That’s good, says Robert J. Marks, author of Non-Computable You: “Images generated by AI should be no more copyrightable than Google search engine results.”