On the Intelligence of AI Graphic Software
You Still Need to Learn to Draw
Text and inages by Glen Bledsoe, 2024 July 1st. (Visit Glen on Flickr)
Two months passed following Stable Diffusion’s initial release in August of 2022 before I decided to try it. Artificial Intelligent picture-making was a rabbit hole I wasn’t sure I wanted to go down. I’d spent years of formal art training—studying life drawing, painting, print-making, human anatomy, learning how to grind my own paint and reading the stories of artists throughout history. I’d always been a reasonably good draughtsman, but learning to translate that skill to the brush took years of work. With my introduction to computers in 1989 I became deeply interested in 3D generating software. These images often needed extensive reworking to fix obvious flaws, and so I learned to use software graphic tools like Painter and Photoshop. With these new skills I explored photography: digital, film, infrared, pinhole, compositing, portrait retouching, landscape and more. AI software promised yet another path to follow. Did I really want to go there?
After creating a couple of images using Stable Diffusion, it immediately became apparent that the news media had overstated AI’s capabilities. “Just type in words and a perfect image magically appears.” That was not my experience. Not at all. The images were often irrational, poorly composed, in need of contrast and color adjustments, and in many instances required extensive repainting to repair major flaws when they could be repaired at all. Dozens of renderings were usually needed to produce a single image, and I often needed to combine the best parts of different renders. The skills I’d acquired from retouching 3D images, portraiture and compositing were very useful. Everything I’d ever learned about painting and art history came back to me as well. Although AI required far more post-processing than photographs, I actually enjoyed the work that took my AI images to completion. Even more importantly the images were finished much more quickly than painting by hand. Which meant I could turn more ideas into pictures.
Over time I found that no one AI software produced images that I wanted. I have had to copy out sections and re-render them in a variety of different Stable Diffusion installs and models. In the end it takes my judgment, my decisions, to determine what needs to be adjusted, what needs to be re-rendered, which selected sections to composite, what needs to be repainted to create usable images. So I asked myself where is the “intelligence” in that? Me or the machine?
Image#1 If AI were truly intelligent would it produce an image that looks like this?
It became clear to me that attributing intelligence to AI software is like a ventriloquist attributing intelligence to his dummy. The dummy presents the appearance of intelligence, of wit, of charm, but in fact those qualities are the product of the ventriloquist, not the inanimate dummy. The ventriloquist brings the dummy “to life.” The same is true with AI software. I believe we’ve been oversold on artificial intelligence. AI graphic software are not “intelligent.” They are an aid—a useful aid which accelerates the production of images. Something to generate useful bits to mold, to improvise upon, but they are not intelligent. And their usability is not always obvious.
The case can be made that AI has no understanding or knowledge of the human body. Fingers! Is there anything more basic than understanding how many fingers a human has? AI has no idea. AI has no sentient insight into expression or color or shape or contrast. No meaning or narration to convey. No imagination. AI just makes predictions. The subtext of all images it provides reads: here is something you will probably like. And then we have to make sense of it and give it meaning. It requires a human mind to take the graphic predictions and select, manipulate, and assemble them into a coherent image to communicate ideas or beauty.
If this were not so then all AI images would look alike and they don’t. Clearly there are different levels of achievement or accomplishments in AI. The difference is not the software, but the person operating the software. AI is just a tool or instrument with which artists create. Put another way, does the beauty of Joshua Bell’s music come from his 300 year old ‘Gibson ex Huberman’ Stradivarius violin or his fingers? It should come as no surprise that without Bell’s fingers and manipulations the violin is perfectly silent. If I tried to play his violin the results would devolve into a form of torture.
In the days when I ground my own pigments to make egg tempera paint I had favorite pigments and brushes that made my paintings possible. But however much they helped me to make the pictures I would never credit them for the creation of my work. I did that. I was the artist. My mind. My sensibilities. My eye. My imagination.
While I choose a hands-on approach to making adjustments and corrections to my AI images because of my traditional media training others choose to use “words only” to fine tune their images. I’m not a software engineer, but in either case clearly a human is at the controls.
When we prostrate ourselves at the feet of AI software granting them artistic abilities that we in fact imbue them with, when we abdicate our role as artist and creator we damage the acceptance of our work and the work of others who use AI as a tool. For our work to be taken seriously we need assert that we are in charge—not some fictional machine intelligence. When we claim “AI made my picture” we create fear in the public perception who already fears an overwhelming onslaught of change brought about by technology and social evolution. We don’t want to flame that hysteria.
Some AI image makers may not agree with my conclusions, and I don’t mean to criticize anyone’s artistic path. I respect everyone’s experience and conclusions. Even so I hope my observations will cause all of us to reflect on what we do as AI artists. Each of us belongs to a long tradition of skilled people that leads back to the days of cave dwellers who had a special way of expressing their ideas, a skill at manipulating materials to produce visions that the eye would otherwise miss. We need to step up and take our place in history.
Image#2 A stage in manual processing the AI result in Image#1 (look at the fingers)
Image#3 The final result
