A Million Little Pieces Of My Mind

As mentioned elsewhere in these pages, I had access to some old family photographs a few months ago; I hastily scanned them with my cellphone, and have been spending the past few months doing my best to restore them, using artificial intelligence as a tool to assist. And I decided to share with you my journey in restoring one particular photo, one taken in 1927 of two of my aunts and my dad's first wife.

This is the original I had to work with: A faded print from 1927 (April 24th, to be precise) of my aunts Lou and Gene with their sister-in-law, my dad's first wife, Lena. To be honest, none of the three look happy to be there! But my task is to use AI (artificial intelligence, if you've been living under a rock) to try and restore it as much as possible.

The first step was to load the photo into Zoner Photo Pro to spot it: That is, to remove little specks of dust from wherever they might be.

The next step was to load the spotted image into Fotor, which has an image-to-image mode. You specify how much similarity you want between your original and the product, as well as in what style you wish to go. Here is the image after whatever processing happened with it set to 100% similarity.

Here's 95% similarity. Notice the features are slightly less skewed.

At 85%, the image looks less like a photograph and more like an illustration inspired by a photograph.

At 80%, my relatives have become cartoon characters. Obviously, I could have stopped here, since for my purposes somewhere between 100 and 95% would be most effective. But now I was curious!

At 70%, they start to look like 1920s biker chicks.

At 50% even the fact that the original was monochrome (black and white) starts to get lost in the rendering. But ya know, I think that they would have recognized themselves in this rendering if they'd had a chance to see it.

At 30% similarity, we're just going crazy. It's the same basic composition—and yet it's completely different, and something definitely out of Ron DeSantis' nightmares.

At 20% and with a different AI model, they ended up zombified!

And at 15%…

Finally, at just 10%, everything is different. It's kind of an interesting illustration, and I might even be able to use it as one sometime—but a restored photo, it is not.

Fotor has several trained AI models that produce different results. Basically, if you train an AI on thousands of pictures from the Pokémon universe's Sunnyshore City, you get a result like this:

And, of course, there's also a Barbie-style AI.

B…U…T… as much fun as that is, it isn't helping us restore the photo. SO, I decided to try using the AI's ability to work with my own prompts instead of a pre-trained model. And at 90% similarity, we have something that does look like a photo. The faces are so prettified, though, that none of them actually resemble the women they are supposed to be. (Sort of like if you got Margot Robbie to play Tilda Swinton in a biopic.)

But at 95% similarity, I got a useable result! While the pained expressions have been lost, I don't think the women would have objected to that. And at the same time I think all three would recognize themselves. (Though, to be sure, I am less certain of Lena since I never knew her; she died before I was born.)

And now for the colorization! (Done by Colorize AI on my phone.)

I have no idea if these colors are reasonable for the 1920s dresses. Aunt Gene's dress appears to be made of satin or some other shiny cloth, so the iridescent look on it isn't completely unreasonable. I could make it a solid color, but I don't actually know, of my own knowledge, what color would be more authentic. So, for now, at least, I am going to leave Aunt Lou, my father's first wife Lena, and Aunt Gene, looking like this.