Learn how to identify an art style for possible reproduction by the AI.
#Tip 17 – How to identify an art style
static sable
Want to know what art style is used in some piece of art that you like? Unfortunately, there is no easy answer. In most cases your best bet is to look at the style directories listed in #resources message to try to find something like what you want. The external guides have hundreds of tested prompts that work with Dalle-2. Selecting a style from those guides is easy and reliable.
When our volunteers are asked to visually inspect art and make suggestions, intuitive approaches often result in suggestions that don’t work, causing you to waste credits. A human can look at something and think the style is obvious, but reproducing it in a reliable manner with the AI is tricky. The only sure way to tell what a prompt will do is to test it with the subject you want. The volunteers want to be helpful, but cannot do testing for you.
Systematic style testing can begin with a “subject-free” test such as “a painting by Vermeer”.
The subject matter and general style confirm that there is training data for this art style. But styles that are in the training data may not work with all subjects. Every word in a prompt affects the results, so the art style can be washed out by other things in the request.
Even if you know the style you want, you may not be able to get it to render if the subject matter is incongruent with the style. Look what happens when we ask for “a puzzled robot is looking at a painting. painting by vermeer.”
Nothing about these images looks much like Vermeer. The robot is a 3D figure, and the art it is examining looks modern. That is because the subject “robot” is overwhelming the art style cue “painting by Vermeer”.
Replicating the art style of works produced by other AI systems is particularly hard because the generation algorithms and training data differ.
If you want to do exploratory research on a image, you can use a type of AI called a “CLIP interrogator” to guess what art cues may be worth trying. These style interrogators have serious limitations and often fail to give accurate analysis, but can be useful. They can look at an image you upload and tell you what they think might be a good text prompt to create similar images using CLIP guided diffusion.
One interrogator that we know of is at
https://colab.research.google.com/github/pharmapsychotic/clip-interrogator/blob/main/clip_interrogator.ipynb
The interrogator is not very user-friendly because it assumes you know how notebooks of this type work. Here is a simple example, testing how well Version 1 of the program works. Version 2 is optimized for Stable Diffusion prompts. Lets use Version 1 to analyze the image we produced using the prompt given above. The notebook has a set of steps that you must execute. The first step is to "Check GPU" by clicking the small arrow.
After that step finishes you will see the GPU listed. Next click the arrow to run the setup. You will also want to include the URL for the image you want to analyze. You can just use the default settings for everything else.
After the setup step finishes, load the URL you want to analyze into the Image specification and then click the arrow button.
The AI will try to figure out what prompt may be able to reproduce the image.
The default settings will show you results from three different AI models.
The summary prompt suggestion is: "a yellow robot next to a painting of a man with a record player, a surrealist painting by Mark Ryden, featured on cg society, les automatistes, academic art, skeuomorphic, dystopian art"
The only problem with this is that we know that the original prompt actually asked for a painting by Vermeer. Just for fun, let's try running the prompt the AI suggested to see what we get.
Interesting, but not a very close match for the original. But it can be fun to look at the smaller parts of the prompt to discover art style cues that may be new to you. Systematic testing of those cues may give you some useful tools for your work.
As you see, identifying art styles isn't easy, and even when you know a style the AI may not be able to reproduce it. Take a look in #📚┃resources and glance through the directories of testing styles to find good ideas for your work!
static sable
For comparison, here are the same steps using Version 2 of the analyzer. It has the same three steps as Version 1.
Here is the suggested prompt to recreate the image: "a statue of a man sitting in front of a painting, a painting, by Giorgio de Chirico, cgsociety, les automatistes, cute robot wooden, mit technology review, interconnected human lifeforms, yellow ochre, artist arata yokoyama, official print, eyes open in wonder". As with Version 1, it has little to do with our actual starting prompt. When we submit that prompt back to DALLE-2, this is what we get:
