Last week, I was wondering how the files coming from a Frontier or Noritsu scan looked so much better than those I was getting when simply digitizing the negatives with a mirrorless camera. It’s because those big lab scanners perform all sorts of adjustments to the images.
But why couldn’t AI do better than a Fujifilm Frontier SP-3000 scanner launched 20 years ago?
Let’s start by dispelling a few common misconceptions. AI models are not truly intelligent in the human sense. They are, at their core, extraordinarily sophisticated statistical models trained on vast amounts of data. Probability underpins every stage of their operation: it is used during training, to infer the user’s intent from a prompt, and to generate the most likely response.
AI models are not infallible. They simply generate the response that appears most probable given the prompt and everything they learned during training. Most of the time, that probability leads to a correct answer. Sometimes, however, the model misinterprets the request or produces information that sounds entirely plausible but is actually incorrect—a phenomenon known as a hallucination.
ChatGPT will tell you that it cannot exactly reproduce the look of Frontier or Noritsu scans. First, the image-processing algorithms used by these scanners are proprietary. Second, the scanners have access to the physical negative, and analyze information that is no longer available once the image has been digitized. That additional information allows them to produce a more refined rendering.
But the real reason why AI can’t totally emulate a Frontier or a Noritsu scanners is that those machines process your images using a well defined algorithm, and deliver predictable and consistent results (it’s a deterministic process).
As opposed to AI which is going to use probabilities to try and guess how it should render the image, leading to slightly different results for each image of a film roll.


In my tests, the color images processed by ChatGPT went from pleasant enough to totally off. AI missed the color of the patio furniture in the example above.
Black-and-white film is a much simpler problem to solve. Without an orange mask or color information to reconstruct, there is far less uncertainty. That makes it considerably easier for an AI model to produce a convincing enhancement.

My tests
I uploaded a few Black and White JPEGs files in ChatGPT – the digitized images captured by my mirrorless camera and the JJC Digital adapter:
- When the images captured with the Argus M5600D were already technically sound—well exposed and sharp enough—ChatGPT advised against performing any AI enhancement on its own. Instead, it recommended creating a custom Lightroom CC preset. I followed its advice, built a “Kentmere 400” preset using the suggested settings, and was pleased with the results.
- Other images, such as those taken with the Kodak Ektar H35, presented a much greater challenge. Many were noticeably soft, blurred or out of focus. I used the picture of Lake Clara Meer as my benchmark. ChatGPT produced a dramatically improved version and again claimed that the same look could be achieved with a Lightroom preset. I followed its recommendations, but the results fell well short of what I was seeing on screen. Even after several rounds of back-and-forth, I was still unable to reproduce in Lightroom the rendering that ChatGPT had generated.

In the end, ChatGPT conceded that, besides all the Lightroom wizardry it had prescribed, it was also performing a number of “optical compensations” that it hadn’t initially mentioned. These were supposedly designed to “carefully recover apparent sharpness and local contrast from a weak lens without introducing halos or crunchy artifacts.”
Nice. But that raises an obvious question. A negative cannot magically record detail that never reached it, so where was the apparent sharpness coming from? Was ChatGPT making educated guesses based on the information still present in the image? Or was it drawing on patterns learned from millions of similar photographs, and quietly generating a new image that only looked like mine?

That left me wondering whether the so-called optical compensations were really compensations. Was ChatGPT recovering detail that was genuinely present in the image, or replacing it with a highly plausible reconstruction?
The Highly Plausible Reconstruction at work
A good thing with ChatGPT is that – if you really insist – it will explain what it has done to the image. Let’s look at this picture of two ladies chatting on a rock in the middle the Sweetwater Creek. It was taken with an autofocus film point and shoot (the Argus M5600d) in the middle of the morning on a sunny day, but the image is blurred, and frankly, unsalvageable, even with the most aggressive Lightroom settings.

I asked ChatGPT to recover it, and to explain what it had done:
“The AI performed a combination of image restoration and content reconstruction.“
- it detected what was probably motion blur,
- It estimated what a sharp image should look like:
For objects whose appearance is well known—such as:- leaves
- rocks
- rippling water
- tree bark
- human hair
- clothing,
- it predicted what those details most likely were
- It preserved the composition
- It increased local contrast
Impressive.
Sometimes, ChatGPT’s intervention goes farther than restoration and reconstruction – it simply composes a totally new image.
There are hundreds of thousands of pictures of the Atlanta midtown skyline taken from Piedmont Park – the shores of the lake are one of the locations where scores of aspiring pro photographers bring their clients for a family shot. It’s likely that the AI models have been trained with a few of them. And ChatGPT sort of remembers your most recent requests.
So, when I prompted that “I’d like a photo of Lake Clara Mere in Piedmont Park taken from the promontory facing midtown. In B&W“, it CREATED an image roughly similar to mine, but with a slightly different angle of view, and without the crane.

As a conclusion
As soon as it’s digitized and becomes a series of zeros and ones in a file, an image can be algorithmically enhanced. The negative scanners used by the pro photo labs have been enhancing our pictures for 30 years now, but AI tools bring the concept of “enhancement” to a totally new level.
ChatGPT and similar AI tools can guide you through the process of inverting and enhancing your negatives in applications such as Lightroom, often by recommending a set of custom presets.
But if they believe that your image can’t be rescued in Lightroom, the AI tools will also perform “content reconstruction”.
That blurs a boundary that photography has traditionally taken for granted. When we admire a photograph, are we looking at the scene that the camera actually recorded, or at a highly plausible reconstruction produced by a model trained on millions of images?
I was half serious when I asked ChatGPT if one of the reconstructed images it had just made available was mine, or its creation.
It’s 50/50, it replied. Before elaborating:
- the creative intent – the decision to press shutter release, the composition, the lighting, the interaction with a living subject belong to the photographer,
- and the rest – the micro-texture in the stone, the edge sharpness, the cloth folds, the subtle skin texture, the tiny shadows, high-frequency details – are clearly an AI reconstruction.
Ultimately, it is the quality of the original image that will make the difference. A well-exposed, sharply focused photograph may need little more than conventional slider adjustments. The AI will suggest how the photographer could enhance the image in a photo editor, and because the final image will be generated by Lightroom, the original EXIF data attached to the image will not be altered.
A soft, noisy or poorly exposed image will invite the model to generate a far more substantial reconstruction. The resulting image will be clearly authored by the AI model, and the file will not include any of the original EXIF data.
Two images enhanced in Lightroom with the Presets recommended by ChatGPT for the Kentmere 400 film stock


A blurry image saved by ChatGPT, and further refined in Lightroom

More AI-assisted images in this Flickr group: AI Assisted Film Photography










































































































































