Bringing Mum & Dad's Photos Back to Life
Not every job is a broken laptop or a printer that won't print! A little while ago a client asked me to help her restore and colourise some of her old family photographs, and in all honesty it was one of the most rewarding visits I've done in a long time.
So the first job was finding an AI that could actually do it. You'd think that would be simple? I tried Siri, Claude and Gemini, and none of them could do colour restoration. They're all clever in their own ways, but ask any of them to take a black & white photo and give it back to you in colour and you get nowhere. Then on to ChatGPT, upload the photo, ask it to colourise it, and back comes a colour version with lovely, believable period colours. Bingo!
From there I put together a simple workflow for the client, step by step, so she could carry on the work herself long after I'd left. She was amazed at the results, and so thankful.
But it also got me thinking. I lost my parents in 2023, and I've got a wealth of photographs, both colour and black & white. If I was going to do this for them, I wanted the best possible result, not just "that'll do".
And that's when I started looking closer at what ChatGPT was actually doing. It doesn't colour your photo, it redraws it. The colours are fantastic, but faces get smoothed and reshaped, and sometimes it adds things that were never there. On a photo of Mum on the beach in 1958, ChatGPT very helpfully added a sunbather in the background who simply didn't exist! For a client's photo that might be fine, but for my Mum & Dad it has to be them, not ChatGPT's idea of them.
So I sat down with Claude, the AI from Anthropic, running on my Mac. Claude can't colour a photo by itself, but it can write and run the code to do it, so we downloaded a couple of open source models to run locally. GFPGAN (a face restoration model) for repairing faces, and DDColor (an open source colourising model) for adding colour. Everything ran on the Mac's own graphics chip, and nothing was uploaded anywhere other than what I'd already given ChatGPT, which matters a lot when it's your family, or a client's.
Did the open source models beat ChatGPT? In a word, no. DDColor left necks grey, couldn't turn a white sky into a blue one, and the colours were muddy. GFPGAN is very good at sharpening up a face, but go too far and it starts to look like someone else, so we only ever used it gently on Mum's face, blended 60/40 with the original so she still looked like Mum.
The real breakthrough was using the best of both. The original scan is the truth, every face, every wrinkle, every stitch of knitwear. ChatGPT is brilliant at colour. So Claude lines ChatGPT's version up with the original scan, takes only the colour from it and lays that over the real photo. You get the real faces and detail with ChatGPT's colours on top. It also cleans up dust and scratches, but very lightly, as heavier cleaning wiped out the texture in the knitwear and the detail in the hands, and I spotted it straight away.
Then came the fun part, a game of spot the difference with every photo ChatGPT gave back. Claude puts the original, ChatGPT's version and the restoration side by side and checks the faces, hands, fabric and anything ChatGPT might have changed. And it changes a lot more than you'd think!
Take this one of Dad and my eldest brother. Dad's holding a monkey, and the monkey is wearing a knitted jumper (don't ask!). ChatGPT turned it into a furry brown teddy with no jumper at all. It also turned the road on the right into grass, and shortened my brother's trousers to show bare ankles that weren't there.
Next came Dad on his motorbike. ChatGPT changed the number plate from YNN 739 to TNN 739, and invented a colourful red, yellow & blue badge on the side of the bike. It's a Gilera, so I found the Gilera logo and sent it to Claude, which coloured the tiny badge in the photo properly, black, white and that red diamond. It's only a few pixels across, but I know it's right.
Then on to Mum & Dad's wedding. ChatGPT gave the bridesmaids blue and pink flower crowns, when in the photo they're as white as Mum's own flower. It also removed the hymn board on the pillar showing hymn 126, which is back where it belongs, and Claude cleaned off thousands of tiny white specks from the dark suits and the church without touching the numbers on that board.
Cutting the cake was another good one. ChatGPT swapped the real decorations on the cake for green leaves and silver horseshoes, when the photo clearly shows dark horseshoes and what looks like a black cat, so they're all kept from the original. It also turned the flowers in the vase into pink roses. We think they were freesias, so the pink stays, but the shapes are the real flowers. And the confetti in Dad's hair? Every bit of it is still there 😁
The last one was different again, my grandparents at dinner. This one started off in colour, but the print had faded badly with a heavy haze across it. ChatGPT made it sharp and bright, but in doing so it redrew their faces, and it was easy to see things had changed. So instead Claude used ChatGPT's version as a guide to lift the haze, keeping all the detail from the actual print, and checked ChatGPT's colours against what was left of the colour in the original. The napkin, for example, was plain grey, not the green ChatGPT decided on. It's softer than ChatGPT's version, because the print itself is soft, but I have to stay true to the photograph, and that's them.
Along the way Claude has kept notes on everything that worked and everything that didn't, so each new photo starts from what we've already learnt, and I've a lot of photos still to go. Is it perfect? No, but every face in these photos is the real face, and that's what matters to me.
If you've got old family photos sat in a drawer that you'd love to see in colour, or just want a hand getting started like my client, get in touch, as always we're happy to help.
Comments