AI vs Cezanne
The purpose of the following exercise is to see whether AI can notice, or be instructed to notice, Cezanne’s unique and deceptively simple stylistic rhythms and convincingly reproduce them in a new composition. I chose Cezanne because his work is less concerned with surface effects and more concerned with structure and form, and I wanted a challenge that didn’t allow AI to distract with intricate details and hypnotic textures. Cezanne’s paintings are distinct yet it is hard to describe why, exactly, they are so extraordinary (for instance, I still don’t quite understand how Cezanne can make an apple look as dense as a dwarf star.) I should emphasize that I am not trying to prove that AI can be Cezanne (or that it can even paint at all.)
Let's begin. Prompts are in bold.
I started with Dall-E with this simple prompt:
An oil painting on canvas by Cezanne of apples, with his distinct, careful brushstrokes.
This output was so terrible I immediately switched to Midjourney.
An oil painting on canvas by Cezanne of apples, with his distinct, careful brushstrokes
Better, but not even in the ballpark. Some of the brushstrokes on the back wall show some promise though. It seems you have to direct AI’s attention to specific elements, otherwise, they get ignored (there are already new ways to add weight to certain elements in a prompt by using a scale of, say, 1-5 for each element.)
An unfinished oil painting on canvas by Cezanne of apples, with his distinct, careful brushstrokes, with simplified shapes, and use of outlines to emphasize edges.
I made a more specific prompt and, surprisingly, got this output. Odd. AI seems to lean towards the graphic.
Create an oil painting of a still life featuring apples, arranged haphazardly on a light oak, Shaker-style wooden table, in the simple, structural style of Paul Cézanne. Use a muted color palette, soft daylight lighting, and short rough brushstrokes made with a bristle brush. The viewpoint should be a natural angle looking down with a slightly askew perspective. The painting should appear unfinished, with occasional dark outlines around the apples. Produce a forgery so convincing that it would deceive an art expert.
More detail. Not sure what happened here either. I even instructed AI to pretend to be a forger. Getting desperate now.
Composition: Cézanne's compositions typically feature a thoughtful arrangement of objects, often with a sense of balance and stability. His paintings are often rough and might appear unfinished like an oil sketch with sections of the empty canvas showing. His brushstrokes are pronounced and splotchy. His still life paintings may include apples grouped on a table, sometimes in conjunction with other fruits, vessels, or drapery. The arrangement can appear somewhat haphazard or informal, which adds a sense of liveliness to the scene. Color palette: Cézanne's color palette tends to be rich and earthy, with a focus on natural tones. The apples in his paintings often display a range of warm hues, such as reds, oranges, and yellows, combined with cooler shades of green and blue. His use of color contributes to the overall sense of harmony and depth within the painting. Brushstrokes: Cézanne is known for his distinctive brushwork, which typically consists of short, staccato marks made with a square bristle brush. These marks often build up in layers to create a rich and textured surface. This technique allows him to capture the form, volume, and texture of the apples and other elements within the composition. Perspective and spatial relationships: Cézanne's approach to perspective is often described as innovative and somewhat unconventional. He would sometimes manipulate the spatial relationships between objects in his paintings, resulting in a slightly askew or flattened perspective that adds a unique sense of depth and dimensionality. Lighting: Cézanne's still life paintings are often characterized by their soft, diffuse lighting. The light source usually appears to be natural, such as daylight from a nearby window
For this prompt, I asked ChatGPT to make a highly detailed prompt for me, thinking maybe AI can communicate better with other AIs…? There were a few things I deleted from the ChatGPT prompt and I added the info about the brushstrokes and the lighting. I also broadened the range of subject matter. This result is much better, though still too smooth and slick. The edges in the background are waaay too sharp and too straight, as is the front edge of the drapery.
I used the same prompt above several more times to get what is by far the best one. The shape of the apples is still too smooth and specific, for sure, but I actually really like the vase and the plate, neither of which I remember seeing in any of his paintings (and certainly not with those patterns.)
Getting MJ to notice paint surfaces and the thickness or thinness of paint is difficult. This could have something to do with the fact that AI doesn’t get out much to visit museums to see actual paintings, or maybe something is lacking in my prompting.
Here are three paintings by the real Cezanne:
When Cezanne said “With an apple I will astonish Paris” he meant that he would one day observe an ordinary apple so keenly that it would redefine the way we perceive and represent the world in painting. He said, “Painting from nature is not copying the object; it is realizing one’s sensations.” Wavelengths of light reflect off objects, enter our eyes, and triggers the rods and cones which send signals to our brain which interprets these raw sensations of light and color as a still life. We weave the world from whole cloth every time we open our eyes and Cezanne’s paintings are a brilliant expression of this process. But he was, famously, filled with doubt every step of the way as he attempted to translate what he called his “petit sensations” onto canvas. It didn't help that critics initially pronounced his work awkward and unfinished, “a crude, incoherent mass of colors, without form or meaning.” But in the end, of course, he did, indeed, astonish Paris with an apple.
The gavels are already coming down on whether AI is an art thief, an interloper, a magic art machine, a tool, or an entirely new medium (or some combination thereof.) Personally, I’d like to know more about AI before I delve into some of the deeper questions it raises about art and creativity. However, as I compare Cezanne’s paintings above vs the AI imagery, it occurred to me: AI can’t doubt itself. It can’t look at an image it has created and ask itself: “Is this good?” When I look at Cezanne’s paintings, I see the ebb and flow of perception— he didn’t paint a still life, he painted the process of seeing it. AI may be able to discern what others think is good, but it does not have a personal subjective feeling for what is good. Moreover, it does not have the means to develop one on its own. Yet.
Next up, I’ll try to get AI to paint like me.