Garden Railway Forum

I've been playing around with getting ChatGPT to colourise some Southwold Railway photos.

Eg
Rik

I had to tell it that the loco should be GER blue

And then tell it to remove the GER lettering

Interesting it's also removed the caption from the original postcard


Rik

I wanted to change the colour of the loco in the shed but I've now exceeded my daily limit for the freebie processing

Rik

Looks good Rik. I did some playing with colourising back when I built the Chingford contractors loco, it wasn't bad but I think things have moved on since then. I actually used the colourising feature of "Find My Past" for that, which was the only easily available one at the time.

I notice on your first set of pictures that it has actually coloured the "green" spectacle in blue, which is of course correct but it could easily have missed that one!

Looks good Rik, and that charming blue engine sure does look familiar in color

I have occasionally seen anachronisms in AI-colorized works, probably due to modern biases in the training data. Stuff like coloring pants as blue jeans before they were commonly available. Just something to watch out for!

philipy wrote: Tue Jul 21, 2026 4:15 pm I notice on your first set of pictures that it has actually coloured the "green" spectacle in blue, which is of course correct but it could easily have missed that one!

Yes, I spotted that.

But I've also just noticed that the spectacle on the far side of the signal ought to be showing green to the train. Clearly, some maintenance is needed. I must have looked at his picture dozens of times and never spotted it.

Rik

It let me have another go this morning so I asked it to colour the loco in the shed. Interesting result.
However, when I asked it to colour the face of the driver, it told me I'd reached my limit again and will have to wait five hours.

Rik

ge_rik wrote: Wed Jul 22, 2026 7:47 am
However, when I asked it to colour the face of the driver, it told me I'd reached my limit again and will have to wait five hours.

I had that problem with Google Gemini when we first got involved with creating figures with Gemini and Meshy. I think its something to do with how they work, so if you ask it a question, or to do something, on a long or complicated thread, it has to reconstruct the entire sequence from the beginning to add to it. It may help, once you get back in, to save that picture, then start a new chat and load the picture back in and simply ask it to colourise the drivers face. No guarantees but doing that has helped me in the past.

philipy wrote: Wed Jul 22, 2026 8:01 amI think its something to do with how they work, so if you ask it a question, or to do something, on a long or complicated thread, it has to reconstruct the entire sequence from the beginning to add to it.

It's simply rationing your (free) use of the AI engine. If you were paying for access you'd continue and be using up your credit instead.

I'm not entirely convinced about recolouring old photos, ultimately it's just fantasy colours. Possibly worse is that in future it may be taken as authentic without an AI tag, with the added problem that when AI recolours an image it often makes other changes to the content as well as just colour.

Paul_in_Ricky wrote: Mon Jul 27, 2026 8:59 am
It's simply rationing your (free) use of the AI engine. If you were paying for access you'd continue and be using up your credit instead.

No, that is different.
I had a couple of instances a while back with a chat that went on over several days and got mega long and it was taking forever to respond. When I asked it why it said it was to do with the length and suggested I start a new chat. I asked how it would know where to continue and it gave me the wording for a new chat that would give it the essence of what had gone before. I tried that and the speed went back to what it had been initially. The only downside was that it had forgotten, or not recognised, some of the earlier stuff in the original chat and I did have to remind it of salient points from time to time.

philipy wrote: Mon Jul 27, 2026 9:51 amNo, that is different.

No, you're mixing up why responses get slow from why you get stopped dead and asked to wait hours for the next step. That's rationing.

Quite often 'chats' get bogged down as the AI makes mistakes, you correct/refine it's answer and it starts going round in circles. At that point you're better off just starting a new 'chat' to get the AI to start from scratch again.

Paul_in_Ricky wrote: Mon Jul 27, 2026 8:59 am ...... the added problem [is] that when AI recolours an image it often makes other changes to the content as well as just colour.

I noticed this. In addition to turning the loco around in the shed, it added an extra chimney to the roof of the cab. When I asked it to correct this and also to colour in the face of the driver in the nearest loco, it removed the original driver completely and added a driver which, to my mind, resembles JD Vance ...... It also removed the chimney of the foreground loco and changed the background loco into something weird. It's clearly trying to be too clever.

Rik

It begs the question, when they use AI to target munitions in a war situation, will it actually do as it's told, or will it try to be too clever ......

Rik

ge_rik wrote: Mon Jul 27, 2026 2:57 pmIt's clearly trying to be too clever.

I think the expression you're looking for is AI Slop

It begs the question, when they use AI to target munitions in a war situation, will it actually do as it's told, or will it try to be too clever ......

The sort of dedicated AI used for 'serious' applications is a different beast from what the casual consumer gets to play with.

Paul_in_Ricky wrote: Mon Jul 27, 2026 8:59 am Possibly worse is that in future it may be taken as authentic without an AI tag, with the added problem that when AI recolours an image it often makes other changes to the content as well as just colour.

Taking a moment to sound like the armchair preservationist brigade, but the way some historical societies are run, AI would make it so very easy to lead them astray on livery - or to correct them.

Before I saw this thread today, I was mucking around again with some colourisers last night. Amazingly, I didn't ask the droid to fix the damaged sky, it just did that.

Can't say I know what livery this Fowler wore (or where it worked, the photo is from my inspiration folder for "the big layout") but when I do know how the paintwork should come out, I notice all AIs struggle to get locomotives right.
Certainly a fun tool to play with, and it really brings a scene back to life.

No idea what colour it should be either, but to my eye it looks totally believable. Sky is good too, as you say.

It's remarkable how it figured out the difference between mountains and sky. I struggled to make sense of it. I actually had to refer to the AI image to draw the distinction between them in the original photo.

Rik

ge_rik wrote: Sat Aug 15, 2026 7:47 amIt's remarkable how it figured out the difference between mountains and sky.

Automated masking for photos has been around for quite a long time now. Adobe Lightroom added 'AI' mask creation for skies back in 2021.
It occasionally gets it wrong, but generally it's very good indeed.