ProfiFoto | Interview on Machine Learning Imagery
This is a transcript of the original interview. If you want to read the German original, visit ProfiFoto’s webpage.
Interview: Boris Eldagsen on AI systems
Machine Learning Imagery
22 August 2022
Machine Learning Imagery is the creation of images with Artificial Intelligence based on image uploads or text input. Fotomedia artist Boris Eldagsen – among other things a member of the German Photographic Academy, DFA, where he is responsible for everything digital – talks to ProfiFoto about his experiences with Artificial Intelligence (AI) to create photorealistic images.
ProfiFoto: Boris, you were kicked out as a beta tester for DALL E2, currently the most advanced AI system that can create photorealistic images and artwork from a description in natural language. What went wrong?
Boris Eldagsen: Nothing, I wanted to test the limits of the system, and that was successful. DALL E2 is similarly restrictive as Facebook. Nudity, sex, violence, war are undesirable, users are supposed to create images like “An otter with a pearl earring, painted by Vermeer” or “American football player in the style of Van Gogh”. I wanted to find out if I could use the AI to generate additional imagery for a current project (“Trauma Porn”, in which I and my artist friend Tanvir Taolad use private photos from World War II acquired from German estates). I quickly realised that I would get no result with entries like Wehrmachtsoldat or 2.Weltkrieg and would be warned. But there are workarounds. And it appealed to me to find them out. A tip: Don’t describe actions with verbs, but describe how bodies relate to each other – or describe the emotional result instead of the action.
ProfiFoto: How does the system work for you as a user?
Boris Eldagsen: There are two ways to create images. Firstly, via an image upload. However, these must not be portraits, as DALL E2 currently does not allow the faces of existing people to be processed. Secondly, via voice input, so-called “prompts”. For me, this is the more exciting option, because you have to be able to visualise the image beforehand and then describe it in speech. This can be very simple, see the otter example above, or very detailed, such as “A black and white photo of a fat otter with three paws biting into the nose of the girl with a pearl earring painted by Vermeer.
with Vermeer weeping on his knees, photo in the style of Cartier-Bresson”. I prefer the latter. If you use blocked words, you get a warning.
Both workflows produce four versions each. From here, there are again two ways to continue:
Via “Edit” I can delete parts of the image I am dissatisfied with, and with a modified prompt have it calculated again in four versions. Or I take a generated image that I like and let the AI improvise over the whole image without a prompt and calculate four more variants. The aim of my test was to create images with the AI via voice input that break the rules of the system to such an extent that the AI refuses to execute the “edit” or further “variants”. Then you get a warning. I had about 50 warnings before I was deleted.
What I wanted is for the image making to move away from the “Decisive Moment”, to painting. Meanwhile I know that working with prompts is more like sculpting, with each loop you “chisel” the image finer from the AI.
ProfiFoto: What other AI systems for image creation are currently available? How do they compare?
Boris Eldagsen: There are more and more providers coming onto the market. The best-known alternative is Midjourney, which runs on the Discord server, or Disco Diffusion by Google. DALL E2 is clearly more user-friendly, but I had to wait six weeks from registration to be approved. New is dreamstudio, which aims to be as good as DALL E2, but without censorship.
I have been testing it recently and can say that there are few blocked words in the prompts, the results are far more explicit and dreamstudio does not give warnings either. But it is definitely more expensive. But with dreamstudio, Pandora’s box is open and it won’t be long before AI porn generation reaches the masses. 2023/24 at the latest.
ProfiFoto: You as a creative – do you see AI systems rather positively or negatively?
Boris Eldagsen: Since I don’t earn my money with illustration, stock photography, commercial photography or pornography, I see it positively. It’s another tool, and like Photoshop, it depends on how you use it. But it will very definitely destroy the professions I already mentioned. It’s especially hard for them, because the AI is using the images that are available on the internet and that have been produced by these professions in the past.
ProfiFoto: What improvements would you like to see in such systems?
Boris Eldagsen: In the AI photo area, DALL E2 is very bad with faces and realistic representation of scenes that are not close-ups or portraits. The highest resolution is currently HD. But this was unthinkable a year ago. As soon as these “teething troubles” are overcome, our image production will change radically.
ProfiFoto: How will the possibilities of such AI systems change photography?
Boris Eldagsen: The distinction between “fake” and “real” will become even more difficult. Anyone can generate and circulate a photo with the Pope in a striptease bar. This will lead to massive political and economic damage, as it will naturally be used as a weapon. At some point, no one will assign a “reality content” to a photograph. Photography will become another form of painting. In pornography, real bodies are no longer needed and the production of images is completely left to the cinema of the mind.
ProfiFoto: So the negative aspects clearly outweigh the negative aspects for the perception of photography in society, even more so than when Photoshop was introduced some 30 years ago?
Boris Eldagsen: The effects will be more massive for society as a whole. At the moment, I see more disadvantages than advantages socially, more advantages than disadvantages creatively.
ProfiFoto: As a photographer, you supposedly have the image in your head before you take it. So in the end, is there no difference between “real” and AI-based images?
Boris Eldagsen: Yes, there is, because most photographers react to a world that is in front of the lens. With AI, they have to draw the content from THEMSELVES. This is a radical change, a fundamentally different approach. Artistically, I have been making this shift for a long time, so AI is enriching for me. Anyone interested in the subject: The Verein Berliner Künstler (VBK) is starting a 7-part series of lectures and discussions on the subject of art and artificial intelligence in September, which will provide a good introduction to the subject. I am happy to host the free Facebook livestream for the VBK.