Written by 
Modern AI models may seem magical, but can they replicate the brilliance and the vibe of some of the most recognizable photographs of the 20th century?
Why Did I Attempt This Experiment?
AI models are becoming so powerful it seems you can create almost anything by using the right prompts—so I simply wanted to test if this is true. While similar experiments scare most creatives, photographers are generally safe from the future dominance of AI for one simple reason: they’re documenting reality.
The photographs you’ll see in this article are not only marvels of the medium but also a time capsule of the period they were taken in. An AI model may be able to generate something with a similar vibe, sure, but it’s really ludicrous to expect an AI-generated image to hold any amount of cultural importance.
My tool of choice was Fotor AI, a relatively simple text-to-image generator that generally delivers solid results. That said, it’s no Midjourney and although it’s great for a variety of art styles, its “photography” skills are oftentimes shoddy. This is especially true when trying to prompt the model to match the time period of the original image.
How AI Handles Human Subjects
Dorothea Lange is the first legend to get the AI treatment. “Migrant Mother” is one of the most iconic photos of all time and I wrongly assumed Fotor would be familiar with it.
Yet, this particular model isn’t that good with names and I basically had to describe the image in detail until it caught on to what I had in mind.The results are solid:
Upon closer inspection, the biggest issue is the lighting. In the original, the subject is lit with natural light, whereas AI gave our “mother” modern studio treatment. The image also looks too slick in my opinion, although this isn’t something that a bit of noise and film grain couldn’t solve.
In terms of composition, it’s a bit of a failure. Dorothea Lange carefully composed the shot to tell a story of someone affected by the great depression, with each detail from the crop to lighting contributing to the message. On the other hand, AI still struggles with subtext (we’re all screwed when it enters its teenage angst phase), so none of that translates into the final result.
Fotor also provides the option of using image-to-image, and this yielded closer results (at least in terms of composition):
While these are hard to tell apart, I’m not satisfied with the results because it looks like Fotor slapped on a low-quality beautify filter on the subject. But what about a more complex scene?
For this, we’ll fast forward to the early 70s and try to recreate “Nixon in China” by Oliver F. Atkins. Below is Fotor’s take on it. The first thing you’ll notice is that Fotor apparently doesn’t know who Richard Nixon is (the model would probably explode if I used a “Watergate” prompt).
The gist of the photo is there (including the bewilderment at a piece of food in the chopstick), but the composition and the setting are too different. I would probably not be able to recognize what the AI-generated image is based on if I hadn’t seen the original.
Image-to-image has the same composition but the overall quality got a major downgrade. The faces are distorted, one of the many signs of poor AI-generated images.
How AI Handles Scenes With Inanimate Objects
Let’s start with the masterwork “The Hindenburg Disaster” by Sam Shere. This once-in-a-lifetime shot is super complex, so it would be a good challenge for the AI.
Fotor more or less understood what I asked it to do.
I used the expression ‘more or less’ because the zeppelin is still fully intact despite the fact it’s engulfed in flames. It is also too close to the ground and it was impossible to replicate the composition of the original.
When using the original image as an input, it’s next to impossible to tell the two images apart:
What about a scene that isn’t instantly recognizable? The Cactus Hot Air Balloon by Leon Gimpel fits the bill perfectly.
It’s also one of the rare examples of the early 1900s’ use of color photography, so it’s interesting to see if AI can recreate the unique look. Spoiler alert: it can’t.
The setting (Grand Palais, Paris) is correct, but the composition and the subjects barely resemble the original. The colors aren’t period-correct, either. Using the original image as a prompt yielded much finer results:
There’s just a tiny problem though. It seems that Fotor simply reproduces the original image in a watercolor/pastel art style, despite using negative prompts to stop the model from doing so. Thus, you shouldn’t expect much in terms of enhancement here.
Can AI Recreate Landscapes and Urban Scenes?
Because of his use of orange and yellow B&W filters to play with the contrast and his virtuosic approach to the darkroom, Ansel Adams’ style is instantly recognizable and very tricky to replicate.
So of course, I decided to give Fotor an inferiority complex. While the scene doesn’t look remotely similar, the contrast does give hints of Ansel Adams.
Image-to-image is much better in terms of overall look but the image lost all the finer detail:
I’m a big fan of street photography, but unfortunately, most iconic street photos aren’t yet in the public domain. Thus, the iconic photo “Kissing the War Goodbye” by Victor Jorgensen will do.
Just as you can expect from Fotor at this point, the final result is decent. There are two problems, though.
First, it closely resembles another photograph of the scene taken by Alfred Eisenstadt, and, second, it now looks 100% staged (the original wasn’t).
On the technical side, the lighting resembles modern-day studio lighting and the photo is too contrasty for my liking. Let’s use the original image as a prompt:
This one is a mixed bag. While I love the look of the main subjects, the faces of the people in the background now look straight out of a horror movie—never a good look.
Overall, I’m impressed with some of Fotor’s creations, but the results are too inconsistent for my liking. Midjourney and DALL-E still take the cake when it comes to prompt-based image generation, especially if you want to generate realistic-looking photographs.