A couple of us have experimented with AI photo backgrounds. I wanted to see the strengths and weaknesses of Google Gemini and Microsoft Copilot. My first attempt was with Copilot. It was unable to create or even repair what I requested and eventually crashed. Here's the image. -- Russ
So I asked Gemini to create the same image. It asked more questions at the start and, on its second attempt produced this image. I never asked for a photo of a diorama, nor for a sign or the botched dirt road crossing. It looks as though Gemini found a diorama photo somewhere online and pasted my depot on top.
The next day I asked Gemini to create a rendering of my drawing of the Eureka Mill 7 ton Porter saddle tank. Instead it found and modified a photo of a larger Baldwin loco and gave it a fancy paint job. Nice photo, but nothing like what I requested. Gemini explained it can't create a photo rendering from a plan. Instead it looks for specific features, such as the pilot design and the cylinder location, then modifies an online photo. Disappointing but the picture is nice.
Gemini asked if I wanted to try something else so I uploaded a photo of my scratchbuilt Carter Brothers boxcar and asked it to place it in front of a circa 1890s wood depot. It produced a very nice image ... without my boxcar. Clearly Gemini has potential but ain't ready for prime time.
So I gave Copilot the same instructions. It started with something vaguely resembling what I'd requested so I asked it to make some corrections. The result was reasonably close but when I pointed out a couple of errors it crashed. I had to sign in formally before it set aside enough RAM to produce the final image and it required about three attempts and another crash. The depot looks like an artist's rendering of somebody's model.
I sent the images to my friend and former contributor, Ed Morris. He spent two days creating photos and my next post shows the results.
I'll let Ed describe the fun he had with Gemini:
First image: A nice image, but even Wiley Coyote wouldn't be tricked by that tunnel portal. Interestingly, Gemini flipped the locomotive to a view of the engineer side and it looks correct. I might try and remove the tunnel portal as I otherwise like the image.
Second (look at the rear of the train): Shouldn't water be dripping off the train? Waterproof firebox? The only way to save this image is equip the engine with a snorkel.
Third: Is that a tiny train or are those giant saguaros?
Fourth: I asked Gemini to change the locomotive number to 105 and the road name to Western Pacific. It did it quite well. The font even looks right. Not sure how far the train will get with soggy coal, though.
This morning Ed sent this photo. He says, "I told Gemini to use Thomas Kincade style lighting on all windows and lights. On its own, Gemini added his style to the whole image."
Nice image, though.
Stuart and Chuck seem to have achieved more accurate and less fanciful results with ChatGPT. Maybe, for now, it should be our first choice.
Russ
[/quote]
Quote from: finescalerr on March 21, 2026, 12:40:42 PMStuart and Chuck seem to have achieved more accurate and less fanciful results with ChatGPT
The various AIs tend to have specific strengths and one of ChatGPT's is images. However, it's worth noting that the image generator is distinct from the core. Think of it as the core subcontracting the image generation. You can ask for edits and you'll get them, but other elements can drift as well. Frustrating, but one approach that can work is to ask the core how it would achieve the result you want. It will help you work around the limitations of the image generator.
Sometimes instead of asking for an image or edit it is better to ask 'how would I get the image?' The core then helps you guide the image generator - to various degrees of success..
The last one looks like a home-made Christmas card !
Barney
Number 4 image 3778 the Depot with the guy standing Looks very nice to me excellent
But Chucks and Stuarts look well just outstanding
But really I should not be making comments on a subject I know nothing about what so ever !!
I can only say what looks good to me and my first impressions.
Barney
None of the images I posted is supposed to be impressive. I posted them to show what different AIs are likely to create, including their laughable errors. Lawrence's post is the key to working with AI and should be in boldface type. It is the first thing to know before fooling around with what's currently available. -- Russ
Thats my problem -never read the words just look at the pictures Just like my weekly comic mag
Barney
A good friend has been asking ChatGPT how to do/find/create all kinds of things. He's emailed me a number of things it found for him to send me about model railroad stuff. The most recent was a long article about the Underground Railroad before the Civil War. It didn't mention what scale it was.
Quote from: Bill Gill on March 22, 2026, 06:13:10 AMA good friend has been asking ChatGPT how to do/find/create all kinds of things. He's emailed me a number of things it found for him to send me about model railroad stuff. The most recent was a long article about the Underground Railroad before the Civil War. It didn't mention what scale it was.
;D
In a more philosophical vein:
Inaccurate output is not uncommon and ChatGPT has a prominent disclaimer. However I think such examples miss the point (other than the comedic). These things are tools. Used properly and with care they can produce great results. Like any tool, when used poorly, inappropriately or without skill, the results are rubbish. They are also developing fast, so expecting consistently high performance is unrealistic.
With the figures, the first thing I did was some research and testing to find out if AI could do the job, and which was most appropriate for me. I then developed (and continue to refine) a workflow that works and that incorporates AI in the process. Finally I'm not expecting miracles. I use my knowledge and wants to drive the process and use the AI for the bits that I either cannot do, or that AI does better and faster. Most importantly, I assess the results and choose to accept them, refine them, improve my process or junk them.
Forums like this typically demonstrate one's best work. For some reason, when it comes to AI the opposite often applies. Granted it can be pretty funny, and I've had the odd 3 legged figure and bizarre clothing arrangement.
Chuck and Stuart illustrated that good results are possible. The existence of bad results does not alter that conclusion.
Lawrence, I'll go stand in the corner :) .
But you're right. AI can be another useful tool that can help do things otherwise that couldn't even be considered.
My very limited sense is AI for modeling currently is like 3D printing was initially when some 'real' modelers felt it was just a novelty or 'cheating' or could never equal the craftsmanship of real modelers - and look where that is now!
There has always probably been some of that attitude. Very earlier model railroaders had to fabricate all kinds of items for their layouts. Articles about building a locomotive began with what size brass stock to get rather than modifying a kit.
Don't let the Luddites (good grief, I hope I'm not becoming one of them!) ruin the challenge.
We're using a wide variety of AIs in my architectural practice, and their capabilities vary a lot. It's clear to me that we're starting to see a gulf widen between free versions, and licensed versions. For image generation, we've had the best success with Nano Banana. It's part of the Gemini suite, but Autodesk has embedded it as a rendering option in the Revit software we use for 3d modeling.
And as Russ has experienced, there's a significant learning curve with all of the options...
My friend and former contributor, Ed Morris, has been working with Gemini/Nano Banana a lot over the past few days. While his communication with AI, and hence his photos, have improved the photos all look like layout shots from a Walthers catalog. He can't persuade Gemini to create something more realistic. I experienced the same result. AI may be evolving at an exponential rate and should improve quickly but it's also important to realize that much of what we see or read is the result of its programming and resulting biases. For now I've categorized no-charge AI as a toy that sometimes may produce very good results. -- Russ
The depot is mine, everything else is AI. Microsoft Copilot copied the depot and Google Gemini created the rest. It took a few days, a lot of false starts, and corrections and cleanup in Affinity Photo to produce this image. AI pictures never turn out as you anticipate.
Another example is the street scene. I wanted my Swayze Hotel model (second from left) to be the center of attention. Gemini had other ideas. At least the image required only two attempts.
You never know what AI will do and that's why I currently consider it a toy.
Russ