Dall.E Vaterland

Article main image

A while ago, a piece of artificial-intelligence (AI) software called Dall.E was unveiled. The idea is simple: you provide it with a so-called prompt — a string of words that needs to make sense, and the software will generate an image from it, using a large set of pre-existing images. For a while, there existed a downscaled public and a limited-access full version. But now, access to the full version is possible for anyone interested in it.

You might have seen examples of such images. Parts of the art world went (predictably) nuts over the idea. This, or so the story goes, is the future of image making. I personally am much more skeptical, in particular given that art is a little bit more than merely taking what already exists to assemble it into something new. Consequently, while I find some of the output presented by Dall.E amusing, most of it is hardly any more interesting than the prompts themselves. For example, when I fed the limited version the prompt “Olaf Scholz eating a potato” (Scholz does look like a potato) the outcome looked as if someone attempting to imitate Francis Bacon had painted the scene. Where Dall.E results get more interesting, you typically land in the area of surrealism. This is good, but we’ve had surrealism as an art form for a long time already. There really is no new ground broken here.

Regardless, when I learned that the full version was accessible, I wondered whether it was possible to test the software on a photographic level. By this I mean that I wanted to see whether it was able to create believable images that could operate inside the context of photography I work in, fine-art photography. To that end, I decided I would have Dall.E generate its own versions of the photographs in my book Vaterland.

The idea is simple, but I also needed a few parameters. For each photograph, the prompt would consist of a simple description of the picture. I didn’t set out to necessarily reproduce the photographs in a one-to-one fashion, though. What would be the point of that? Instead, I wanted to get images that corresponded closely enough to the original photographs so that they could replace them in the book.

I realise that this description might sound a little bit vague. Look at it this way: in the book, the photographs exist not only as individual images but in particular as elements of a larger sequence. Consequently, they serve a function inside the larger whole. If you extract a photograph from the sequence, you only get part of the full function in the book. Therefore, for my experiment it was more important that the Dall.E images would get close to the functions of their corresponding images in the book than that they would look exactly like them. That said, if Dall.E handed me something that was very close to the original, I didn’t reject it.

Coming back briefly to what Dall.E does, I don’t want to approach it in a dogmatic fashion. Assuming that its output improves, I am not ruling out future use. However, I would never use it for a project such as Vaterland. Here, and in the follow-up project that I’m working on, discovery plays an important part. This discovery happens on both the photographic level — coming across photographs that I could not have imagined — and my personal, artistic level — gaining insight into aspect that I previously had not thought of.

Furthermore, I believe that I need to make myself vulnerable in my work. However uncomfortable this might be at times, it is from that spot that the most interesting results have arisen in all of my work (whether in my writing or my photography). Using tools such as Dall.E mostly precludes discoveries. Fair enough, as a photographer you don’t need to use AI to avoid discoveries. Many photographers do really well without, illustrating their pre-visualised prompts themselves. As a viewer, critic, and photographer I personally am not interested in such work.

I’m not going to reveal most of the prompts. All I will say is that they were very basic and obvious. Each one described a photograph with a brief sentence. For example, the very first picture in the book is “An empty construction site that is fenced off in front of a number of apartment buildings in Berlin”. That’s it. Because I didn’t want to try to get as close as possible to my own photographs, I didn’t feel the need to use very specific prompts.

Of the four images Dall.E offered me I picked the picture that came closest to what I needed. When two or more pictures worked, I picked the one I liked the best. This obviously is a subjective choice on my part. But a choice had to be made, and in the usual spirit of my work, I wanted to keep it simple.

Once I had made my choice, I treated the Dall.E image as if it were a raw file from my camera. I converted it to b/w and worked it over in Photoshop to make it look like my photographs. The tones of the b/w I’m using in my work are very deliberately chosen to evoke an atmosphere. If the following helps you understand this, feel free to imagine that I added a filter to the Dall.E images (unlike many — most? — photographers, I don’t have a problem with the idea of “filters”).

Once I had started, I noted that performing this processing step also helped bring the images closer together. The colours, saturations, and the way light was treated in the Dall.E images varied widely. Those differences would have made it almost impossible to have them work in sequence. In two or three cases, I also flipped an image horizontally. This was also done to have it work in the sequence it would have to operate in.

The Dall.E interface produces square pictures. My own photographs follow the 4×5 aspect ratio. I only learned that you can change the Dall.E output once I had put everything together. But I actually don’t mind the difference in formats. As I said above, I did not try to produce exactly what I had made (because, again, what’s the point?). Furthermore, as you will see below once I overlaid the Dall.E pictures on top of my own in the book, a part of the originals peeks out, reminding the viewer that there is another picture.

To add the pictures to my book I needed prints. The final step of the process involved getting them at a copy shop (fun fact: I don’t own an inkjet printer). I added these prints to a copy of own book by “tipping them in” (I used a piece of tape). Now, I have two versions of my book: the original and a Dall.E version. In the following, I want to discuss what I learned from the process and show you some examples.


The full article (which includes many examples) is available for subscribers of my Patreon. Please consider signing up for a subscription. In addition to getting access to this piece the existing Patreon archive will become accessible (articles and photobook videos).

Much like journalism, photography criticism involves a huge investment of time and resources. When you become a subscriber, you not only get access to more of my work. You will also help me produce it (including the free content on this site).

Thank you for your support!