Experiments with AI
Is AI up to it!?
May 2026. I was given this photo of US President HG Bush and asked to retrieve the barely visible signature under the photo. The cream surround to the photo had also faded back to a grubby gray- though the definition of head and shoulders was good. I spent several hours retrieving the image and signature which was written with a fountain pen which had bled into the paper. My results are in the centre. Then, for comparison, I put the original image into Gemini AI and asked it to put a very light cream tint on the photo and to retrieve the signature. It took it about a minute!! but the signature and inscription was not the President's.
Before
Final (Without AI)
With AI (below)
On close inspection the AI signature is too different from the original signature to be acceptable to the client. Signatures of famous people cannot be changed to that extent. AI made a good job of the toning and portrait, but not the signature. Not many people would notice, except for the client, and maybe the president. The Gemini image was also sharper than the client wanted and the smile wider.
Small boy restoration
April 2026: A badly deteriorated photo for restoration. Much of the work on the hole in the middle I had to do myself using photo restoration software. But I used AI to create the woollen knit jersey, shorts, remove some facial noise and put the marble effect on the wall behind.
The Original
Half way Through
Done
I used Photoshop and Topaz tools to get the photo half way. From there on I used AI: a piecemeal approach focussing only on certain areas: getting rid of noise on the face, improving the checked gingham shorts, the finely woven knitted jersey and the marbled effect on the wall at the back. Then it was back to me for finishing touch-ups. I was not entirely happy with the face, but where do you stop?
What happens if I use AI only?
AI's - first attempt
AI's - second attempt
When I simply gave Gemini AI the original photo with text prompts to "repair and restore" it did a good job with the floor, walls and lower half of the boy. But the shirt and facial likeness were not there. I changed the text prompt : "Replace the shirt with a woollen sweater with a flying black duck in the center. The duck is flying obliquely to the right". AI not only did that but put (unwanted) curtains behind the boy and a brown tint on the photo. Using text prompts is a skill that often requires multiple attempts and had I persevered I might have reduced the size of the duck, and turned the sweater into a light woollen knit. But the boy's face would still have been much too far from the original. The thumb of his left hand was hooked into his trouser pocket and not hanging by his side. In the end it was probably faster to do it myself, with minimal assistance from AI - which is often the case.