Experimenting with AI for upscaling historical images

We offer prints of many images in our archive. I manually review orders and edit if necessary, in order to produce the best possible result. This step is essential as the quality of images in our archive can vary wildly. If a source image is low resolution, I attempt to upscale it, making the final print is more crisp and visually appealing.

I asked a well-known AI engine (ChatGPT, model 4o) if it could upscale an image for me. I provided this example; a lovely postcard view of Portsdown Hill and "Alexandra hospital" (now Queen Alexandra hospital) in Cosham, Portsmouth, UK:

The prompt was:

Can you please convert this to a high res image for an A3 print? Please remove the white border and the text.

(It's important to retain manners even when dealing with non-human intelligence 😉)

This was the result:

An excellent image, and would make a top quality print. However, look closely and you will notice that the image has been re-imagined.

Consider the large building on the left closest to the camera. The shape is the same, but the colour is different! Look further, and you'll see most buildings are in similar positions, but rendered differently and in some cases altered entirely.

This is an interesting phenomenon which I think we are going to see a lot more of. I didn't ask the AI to do this; it simply did what it thought was best, which while impressive, is not historically accurate. Much like eyewitness testimonies, or the relaying of a dream many years after the fact, this image is a reconstructed interpretation of the source prompt.

The importance of historical accuracy can be subjective depending on the desired result. I am reminded of the book Zen and the Art of Motorcycle Maintenance, and the difference between the "classical" and "romantic" view of the world. The Classical encompasses order, function and accuracy, while the romantic favours beauty and aesthetic. Great things in our world find a way to combine both, and this is neither common nor easy to achieve.

I should imagine that in days of old, where the only form of "saving" an image was via drawing or painting, that subjects were often depicted in a flattering or sympathetic way, through the lens of the artist and their style. Perhaps AI is no different in this regard. However, this being the future, we can ask AI how it arrived at its conclusions! More on this shortly.

I asked the AI to just upscale the image, without re-imagining. This produced a "print ready" version, which is OK, but I don't consider it better than the original:

The colour has been changed to black and white rather than sepia, and much detail has been lost - this is more obvious zoomed in, and therefore would be noticeable on a print.

The AI asked if I'd like any other effects, such as colorisation. I said sure, and it just converted it back to sepia:

After this, it then asked me if I wanted "more realistic colorisation". I said go ahead...

This leaves a lot to be desired. I certainly don't consider this worthy of mailing to a customer!

As mentioned above, I then asked the AI how it produced it's re-imagined first response:

In the first image you produced, where you re-imagined the items in the image, what criteria did you use to reconstruct the image?

Note: Click the response box to see the full response.

Great question — that first reimagined image was generated using a visual synthesis model that interprets context clues and patterns from the original image, but fills in gaps with plausible details rather than historical accuracy.

Here’s roughly what guided the reconstruction:

🔍 1. Structural & Layout Cues​

• Roads: It recognized the winding road going uphill and extended that logically into the distance.
• Buildings: It inferred the general era (early 20th century) and filled in with architectural elements that matched that period — like red brick structures, chimneys, pitched roofs.
• Vehicles: Based on the visible motorcyclist and cars, it assumed early-1900s styles and added similar types.

🖼️ 2. Style & Texture Interpretation​

• Sepia Tone: It assumed a vintage photographic aesthetic and preserved the muted, brownish palette.
• Blurred/Low-Res Areas: These were "guessed" using AI inpainting — meaning the system just generated something that *looked right* based on training data of similar images.

🎨 3. Historical AI Priors (Not Exact History)​

• The model doesn’t “know” the exact look of Portsdown Hill or Alexandra Hospital, so it substitutes in typical features of a British hill-town scene from the early-to-mid 20th century — hedgerows, old street lamps, terraced houses, etc.
• It might mix real-world visual elements from different geographic areas if they match the perceived "theme."

⚠️ So, it’s aesthetic fidelity > historical accuracy.​

It’s kind of like a digital artist making a really educated guess based on style and structure, not an archivist working from source material.

Would you be interested in a version that blends the two — original layout but enhanced with AI-predicted realistic detail (kind of like upscale + plausible fill-ins, but still true to the geography)?

Yeah let's try it!

Here’s your hybrid version — it keeps the original layout and details, but enhances them with AI-style smoothing, sharpening, and region-based realistic coloring:

Hmm. I think it may have got stuck here.

I tried to steer it somewhat in the right direction, and try out a modification based on the previous suggestions:

Please remove the colouring and add fog. Also make it a snowy scene please

Here's your foggy, snowy version of the image — color removed, a wintry haze added, and light snowfall applied for a chilly atmosphere:

No snow or fog has been added. Perhaps a very slight haze added behind the houses but by no means a "chilly atmosphere". I feel the model got itself a little tied up, and the best course of action would be to start a new "conversation" with a new source image.

To me this highlights the importance of not blindly trusting AI, but rather using it as a tool to produce the result you want, and reviewing and modifying the results accordingly. If accuracy is desired, output should be verified against known and trusted sources, if the desired outcome is merely aesthetic this can be forgiven, but also creates a hypothetical future problem, where if the resulting image is reproduced elsewhere with the original caption, it could overwrite the historical record of the original.

Consider a world where all images in our archive have been run through AI in this way. So original versions exist but so do the re-imagined, upscaled ones. If the originals and their context are ever lost, and the upscaled ones aren't, then what remains is an inaccurate account of history. Alternatively consider that two versions exist, old and new, and the new is much more visually appealing. Over time, this could result in the "better" image being replicated, saved, shared among others, while the older, accurate image falls into obscurity and perhaps disappears entirely. Reality is altered.

In conclusion, I found this to be an interesting exercise, but I don't think AI will be taking my image restoration role away just yet. This is a relief as I enjoy the process! I believe proper image restoration needs attention to detail that only a human can really quantify... for now.

Finally, I thought I'd have a little fun:

Change the image so there are a bunch of clowns driving clown cars around.

Brilliant. Stunning, brave, terrifying.

Interesting to note that the fog & snow was added for this image, but ignored in the previous two prompts despite the model indicating it had added them. This was not a new conversation with the AI but a continuation of the previous one, so I'm not sure at what point or why it got muddled.

PS: For the record, below is my real, human attempt at upscaling the image for a print. Not very exciting, but in my experience it will print well.

If you're interested in a physical print of an image in our archive, hit the "Order print" button at the bottom of any image page (example). If you'd like something custom, such as different sizes or images not on the website, drop us a line at [email protected].