AI Image Manipulation

[LASteve]

@Ensign Elliott pointed me in the right direction with this, although I used Copilot AI rather than ChatGPT to generate a snow scene from a picture of my Moorpark layout. Here's the original and the AI versions:






It's not bad at all. The AI image has few anomalies but you wouldn't know if you didn't see the images next to each other. It couldn't read the "Moorpark Junction" sign clearly, so it generated one of its own. It cropped the left side of the picture, maybe because it was out of focus and it didn't recognize the double-track mainline and made one single track instead.

Took about two minutes to generate the new image. Pretty amazing, really.

[ntpntpntp]

This is the problem with some of these AI enhancement tools, they don't just work on your image they selectively replace stuff.   

I noticed this when I gave one of the cheap online photo enhancers a blurry old photo of two of my long missed cats, hoping it would simply do a bit of sharpening. What I got back was a cute photo but clearly were no longer my cats. 
Worse still, the so called "try it for 10p" offer had actually set up an unauthorised recurring monthly payment for £49.99 from my PayPal account, which fortunately I spotted and cancelled after PayPal prompted me to review it!  (I then found several people reported this thievery on TrustPilot).

[LASteve]

@ntpntpntp Send me the picture of the moggies if you like and I'll run it through Photoshop - I've got access to the full-blown Adobe Creative Suite here.

[ntpntpntp]

@LASteve   actually now that you mention it, so do I have access to Adobe Creative Suite.   I currently have in my possession a recently deceased modelling friend's Macs and other devices while we deal with his estate.  He was a school reprographics tech.

[Malc-c]

This A.I stuff is scary.. its so easy to generate images that look so realistic its hard to tell if they are real or not.  Granted you may find errors, but given its "learning" ability it won't be long before those errors get corrected.  Problem is people do believe what they see... an AI generated video of the late queen dancing in what looked like a care home environment still convinced lots of people that she was still alive !

I've not used it for image work, but I have for programming ESP32 devices.  I described what I wanted the project to do, and what sensors to use and it built the program in a matter of minutes.  I then spent around 8 hours getting it to add functions and cosmetically look a little better - some attempts it broke the code, but it then went back and fixed it.  After a days work I had a fully functional project that would have taken me 3 - 6 months and with a human mentor to debug my code.

The image above looks amazing, very life like and something that could be used for a personalised Christmas card.

[Newportnobby]

Using Gemini.Google I took an image posted by @Bob G  and 'winterised' it.
Note the original has a stain top left and it's now gone

[icairns]

In the November 2025 edition of Railway Modeller, Ian Blenkinsop describes how he took a black-and-white photo, colourised it using computer software, and then scaled it to OO gauge and printed it to use on a building for his layout Cambois Bay.

I am interested in trying to colourise some old B&W photos and wondered if NGF members had any recommendations for a suitable software. 

I would also be interested in software that can remove "cracks" and lines on old B&W photos.  My interest is primarily on colourising railway photos but also possibly some old family photos. 

There seems to be many software companies offering colourisation, such as HitPaw, NCH, liftapp, Canva and many, many more.

I would be interested to hear of any real-world experiences. Also, because I am somewhat thrifty, I would be looking to start with free software (not a free trial), if possible. 

Any advice will be much appreciated. 

Ian

[LASteve]

@icairns - I took a B&W photo of a warehouse, loaded it to (free) gemini.google and gave the prompt "colorize this picture". Orginal and resulting AI below:








Took less than a minute to product the render. Pretty impressive, eh?

[icairns]

Took less than a minute to product the render. Pretty impressive, eh?

Yes, very impressive!  I am going to give it a try. 

Any other other ideas still welcome.

Ian

[LASteve]

And one more ...

[port perran]

And one more ...

Did you tell AI which colours to apply?

If notit's done a pretty good job in my opinion.

[LASteve]

Did you tell AI which colours to apply?

If notit's done a pretty good job in my opinion.

Nope, a three word prompt "Colorize this image". I thought the color rendition was pretty amazing both with the loco (the red buffer beam especially), the carriages and the distinction between the brick station building and the corrugated concrete outhouse.
Looking closer at the tender, it looks like it added the British Railways emblem.

[EtchedPixels]

Using Gemini.Google I took an image posted by @Bob G  and 'winterised' it.
Note the original has a stain top left and it's now gone

As has the curve of the track with the vans in the distance.. which is a lovely example of why it's cute for christmas cards and shouldn't ever be used to "enhance" stuff for research or reference.

[EtchedPixels]

Did you tell AI which colours to apply?

If notit's done a pretty good job in my opinion.

For most surfaces an image a little context and some guesses on the shading is enough to accurately colourise it. My brother is totally colourblind (as in he sees only grey shades although he assigns "colours" to them by learning for compatibility with normal humans). He can accurately recolour most of the black and white things we've asked him about just fine. This has been doable fairly effectively before large language models got used to do it 1000 times less efficiently

My brother was a godsend watching snooker on the black and white tv.