Playing with AI

It's almost impossible these days to ignore discussion about the impact of artificial intelligence and its growing influence on our day to day activities.

Recently, I have been dabbling with some aspects of AI and its uses related to another interest of mine, family history. In particular, I have been using AI software to upgrade old photos and interpret or transcribe old documents. With this experience, I decided to try an experiment on one of my Philip's Creek photos. I have used ChatGPT previously but, more recently, I have started to use Google Gemini.

My backdrops have been painted and I was interested to see if a backdrop could be upgraded to a photographic image. My start point was this photo. 

 

 

 

 

 

 

After copying it into Gemini, I asked the app to replace the backdrop and this photo was its first effort. While impressive, I thought it was too hilly and too heavily timbered for my layout.

 

 

 

 

 

 

So I asked Gemini to change to background to a landscape based on northern NSW and this was the result. Not bad, but with the benefit of hindsight, probably I should have specified the upper Hunter River valley. 

But, for the purposes of the exercise, this was sufficient.

 

 

 

 

Now, I was starting to get a bit over confident but one lesson from my other work needs repeating. AI makes things up if left to its own devices. Two experiences from my family history research bear repeating. One photo, taken in the 1940s, showed a taxi driver standing next to his cab at the entrance of a garage. There was something else in the garage but it was not clear enough to be enhanced, so the AI system added some very modern cardboard cartons in lieu. In a second one, a beach scene, again from the 1940s, it added some extra people not in the original photo. In both situations, this extra content was added without instructions.

The final test was to see if I could add some extra realism to the locomotive. I asked Gemini to add steam and smoke on the locomotive to this photo consistent with a NSWGR 38 class. The photo below is the result. I am quite happy with the outcome but it confirms the current understanding that one can no longer be sure that a photo is an accurate representation of fact.

So, what will I do with this capability? It probably won't involve a mass alteration of existing photos. Rather, I'm hoping to use it to create bespoke backdrops to upgrade the existing painted scenes to a photographic quality. I'm not sure how I'm going to achieve this and it won't happen any time soon as I need to finalise the refurbishment works. It's one of those jobs that are stacking up and it just reinforces the old adage that a layout is never completed!