Pretty hard to restore such a fresco, especially when large parts are simply gone. It takes a long time too. You’d do it for Michelangelo in the Sixtina, you’d probably do it for an original Giotto, but who would do it for the mass of those frescoes?
Maybe this is a job for generative AI? Let’s make an experiment:
Using Photoshop generative fill, it took me about half an hour to come this far. There is no rescue for the dead Saint, especially because Adobe disallows prompting for torso and legs of a dead Saint 😁
I used nine generative fill layers. This sounds tedious, but you just select the area you want to fill, wait 10, 20 seconds, and you get three options. Choose the best, delete the other two and generate another three. See, if you’ve got something better. If yes, keep that and repeat, if not, repeat anyway. When you’re satisfied, select another patch and generate another layer.
Anytime in between, you can change the prompt. All but one of the nine patches in this image were made without prompt. In that case, the AI just tries to come up with something that matches the surroundings.
For the face of the child, I prompted for Giotto’s style. What I got is too small, but it fits pretty well. The face is from the third of four triples I’ve tried. It matched best with the environment and the overall style and mood. Although it did not preserve the original direction of view, I guess most people wouldn’t recognize the fake. It’s a child, just a distracted child, isn’t it?
In reality, as far as we can see in the original, the baby Christ looks up to the face of its elder self. That makes sense, and I am sure that this alone would be a dead giveaway for Art historians. Nothing in Christian iconography is random 😄
In the end, it’s just a tool. You can use it or abuse it. For some uses, it is just the best tool we currently have. In most cases, you could achieve mostly the same results with traditional tools, but it would take much more time.
Don’t try to optimize time though. It really pays to try alternatives, and there are often more aspects than simple visual plausibility.
As always, I could have got better results with a few hours of prompting, and probably trying alternative tools and mixing results. Having done it for fun and the sake of the argument, I am pleased enough though.
PS: It took me more time to write the text than to work on the image. I could have “optimized” with ChatGPT, but doing so would take all joy out of blogging. Another thing I’m spared by not monetizing my blog 😛