Using AI to create and edit textures.

[ossie]

Another try with Chat GPT and a much better (and more exciting) result this time. I'm sure all of us that do reskins have felt that surge of frustration at finding a photo of something we'd love to have but it's on too much of an angle to be of use. This is before & afters of IC 6250 asking GPT to reproduce it in profile. The details of the loco are far from perfect but no doubt if I'd fed some additional shots in for the AI to consider or asked for some tweaks, the end result could have been much better. The possibilities (and particularly for you guys that work in the pre-digital camera era where good pics were harder to come by) really are exciting.


ChatGPT Image Jan 1, 2026, 09_00_32 AM.png

[oper190]

So if anyone is building a route with a bicycle manufacturing plant on it, drop me a line. I've got your shipping needs covered.

Be good for the Mini's also

[LuFeDuTi]

Try before you buy and all that. I thought I'd have a run at Chat GPT before I made any final decisions about a paid option. I asked it to do the following:

I'd like you to remove the shadows from around the end ladders of this autorack and also remove any shadows created by the roof overhang. Keep the image as sharp and detailed as possible and do not reduce the original dimensions of the photograph.



So if anyone is building a route with a bicycle manufacturing plant on it, drop me a line. I've got your shipping needs covered.

A Rare IA New Autoracks Six Ford Cars

[wwhall]

Another try with Chat GPT and a much better (and more exciting) result this time. I'm sure all of us that do reskins have felt that surge of frustration at finding a photo of something we'd love to have but it's on too much of an angle to be of use. This is before & afters of IC 6250 asking GPT to reproduce it in profile. The details of the loco are far from perfect but no doubt if I'd fed some additional shots in for the AI to consider or asked for some tweaks, the end result could have been much better. The possibilities (and particularly for you guys that work in the pre-digital camera era where good pics were harder to come by) really are exciting.

I haven't worked enough with ChatGPT to know how it "thinks," but Gemini does have pretty "intrusive" blocks in its programming to prevent you from trying to use it as a photo editing tool. Another understandable, but inconvenient characteristic that Gemini has is that, if it can't figure something out in a prompt or image, it will "fake it" and add something from another image that it has from who-knows-where into what it renders. For example, putting an EMD SD70M radiator on the back of a GE AC4400 if it can't figure out the details. In AI slang, that is called "hallucinating" something. Also, when you add a subsequent prompt for an already-generated image, Gemini basically starts over and recreates what it thinks that you want. It gets really confused about lettering and numbers. For example, you ask it to render a brick building with a "Grocery" sign painted on the front. It renders it, but in red brick and you wanted yellow brick. So, you make an additional prompt, saying, "Change the brick color to yellow." First, it may not do this and will give you a "I'm just a language model, I can't do that" warning, meaning that it thinks you're just asking it to do a photo edit. If you tell it, "That's almost perfect, but I'm visualizing the brick being yellow," it will probably execute that prompt, BUT the new render may say something like "Graercy" instead of "Grocery." You have walk a fine line to tell Gemini to leave a rendering unchanged, except for the change that you want and having Gemini reject your prompt as being just a photo edit.

Let me illustrate this with an example, this of a non railroad subject, but you'll get the idea. The original prompt that I wrote to get this scene was longer than this post. It got it nearly perfect the first time, but the stars were not quite what I wanted. So, I prompted Gemini with this, "That is almost perfect, except that stars need to be somewhat subdued and that scene darkened slightly. It executed my request, but it altered the lettering on the model no. of the John Deere grader, even though it did not change that part of the image in any other way. Go figure.

Sandhills snow.jpg

[oper190]

I've being trying AI to create images of character for a D&D story line I'm working on, to make sort of a story board. I'm using both GG (free) and ChatGPT (pay) and getting good results, usually cartoonish images rather than photos which is perfect for what I wanted. I'll create several character using photos of real people with the "use this image as a reference and make...." I'll find they render images cartoonish, but sometime photo quality, with the character altered from the photo which is what I wanted, I just wanted to use the photo as a base model. Then I'll try to do another character using the same description and told by both "I can't render an image of a real person." Dumb and Dumber??
I'll then either change to a new Chat string, or signing off and use the other program for a while.
Overall though find it they do a decent job.
Now going to try some of my homemade textures for the sim and see how well it can render them more real like.

[ossie]

I haven't worked enough with ChatGPT to know how it "thinks," but Gemini does have pretty "intrusive" blocks in its programming....

I hear you Wade. I think I said something along the lines of "equal parts frustration & fascination" in response to Pete and that is pretty much how I see AI at the moment. My initial run at my Christmas card for TS.com was a screenshot of one of our trains. Three locos, a crew car and single stacked domestic containers. I asked it to put the train into a typically Australian desert setting and got back a place that looked more like the Gobi Desert or even Mars than anything I've seen here, plus the entire consist was subbed out for double stacked ocean containers, the crew car and 3rd loco were gone and the remaining two locos even had different numbers on them. It was just like you said, as though my input was totally ignored and it went off and found something in the "archives" to fake it with. And yes, the near enough but one final tweak needed step can often set you back quite a distance. If I've got something that is 90%, I often find it easier to put it into a different machine and ask for the adjustment there rather than try to get that last 10% in a current session as that can often get REALLY borked up rather than improved.

I could write an essay on the oddities and inconsistencies I've encountered but I'm sure everyone is experiencing similar. One interesting experience though, one of the kids and I were going to troll my wife by messing with photos of a couple of her favourite footballers. Her favourite player we decided we'd put into the jumper of our most despised club. I put his name in and said replace his football jumper with the one in photo 2. It came back with something along the lines of "I'm sorry I'm not allowed to edit real people" (or words to that effect). Ok, remove the jumper from the guy in photo 1 and replace it with the jumper in 2. Done. I don't know if that's another "I'm a language model" stall or whether they are trying to cover themselves a bit as far as editing real people goes, but if so, our workaround took 2 seconds and we didn't even need to start a new session or replace the photos we'd initially asked it to work on.

In the short term I guess a photo editing tool is exactly what I'm looking for. Once I get to work on the autoracks I might give Gemini a go (I noticed they're offering a free first month) and see how it goes with basic shadow removal etc and hope the process for turning a photo into an accurate profile shot for a reskin and more advanced editing functions becomes a bit more straightforward and smarter as time goes by.

[oper190]

Yah sometimes they do freaky things. Just tried to add weathering to my skin for a boxcar model of my own design. Got a picture of a real box car, not my skin edited. Tried rephrasing, adding that is it is a skin for a 3d model of a railroad boxcar, and still got the same photo back. Got one I'll post,

I call it a GenesisF7. Apparently this engine was in a collision and they gaffed an F7A unit nose to it.
Otherwise a pretty good photo like render.

HRL7 -2ChatGPT Image.png

[pwillard]

Personally, I find the AI assistance with graphics to be the most disturbing/bothersome. This is primarily due to it being a complete crapshoot. I can supply an excellent image and tell the chatbot to "add snow", which it can do... but at the expense of the quality of the underlying brick texture. This is not what I wanted... so I refine my prompt. Six attempts later, all I've really done is waste time on something I could have done myself with a standard bitmap tool. I got lazy and paid the price. Other times, of course, it can do exactly what I was looking for, hence the crapshoot comment.

[wwhall]

Here is the thing--though it won't admit it, though it hints at it if you question it with "Why did you do that?" or "Why did this happen", Gemini is very much a work in progress. The good news/bad news is that it is learning from its mistakes. Gemini, apparently, is also sort of like a dog that you are teaching new tricks. Let's say that you've asked Gemini to provide and image of a modern Australian train in an Australian desert scenes (I like to avoid using the word "typical"--that seems to give Gemini the idea that it has wide latitude to do what it wants), and it renders a perfect image. If you hit the "thumb's up" button, I think that Gemini does two things, a) it has learned something about you (you like Australian desert scenes, for example), and b) it has "learned" right or wrong, "Hmm, this is probably what an Australian desert looks like." In a sense, I believe, hitting that "thumbs-up" button is like giving the dog that you are training a reward treat for performing the trick correctly and the dog will remember what behavior got him/her a treat.

You're running into the Gemini "quirk" that I ran into on some images. I may have posted this earlier (I've slept since then, just like AI), but I'll repeat this part of a response from Gemini when I asked a "why is this happening?" question:

2. Trigger Words: "Remove" vs. "Create"

The specific words you use are tripping the safety wires.

• "Remove," "Edit," "Change," "Fix": These verbs imply modification of an existing file. The Chat Brain is programmed to reject modification requests to prevent deepfakes or copyright infringement.
• "Generate," "Create," "Show me": These verbs imply creation. The Chat Brain is much more likely to allow these without complaint.

3. The "Ghost" in the Machine (Context Confusion)


Because you have a long history of generating similar images (trucks, trains, winter scenes), Gemini sometimes gets confused about whether you want it to look at the previous image and edit it (which it refuses to do) or if you are simply refining your prompt for a new generation (which it can do).

Here is one key that I have used. When you get an image close to what you want, download it and save it to your computer. If Gemini borks what comes after that, then start a new chat, upload that "last good" image and start over from there.

[wwhall]

Yah sometimes they do freaky things. Just tried to add weathering to my skin for a boxcar model of my own design. Got a picture of a real box car, not my skin edited. Tried rephrasing, adding that is it is a skin for a 3d model of a railroad boxcar, and still got the same photo back. Got one I'll post,

I call it a GenesisF7. Apparently this engine was in a collision and they gaffed an F7A unit nose to it.
Otherwise a pretty good photo like render.

Actually, not a bad looking design. In AI jargon, it "hallucinated" the locomotive. One "editing" technique that I've used, and Gemini will usually accept is if I make prompt like this, "This image is nearly perfect, except that I was imagining that the lower headlight below the cab windshields was red, not white. Can you do that?"

[oper190]

Actually, not a bad looking design.

That was actually after a good rendition, but not quite what I wanted, so asked for modification. Yes it improved the photo like quality to what I wanted, but morphed the engine.

As said if you don't get the result you wanted and can't get an expectable improvement, save the best image and start a new chat, using that image as your base. I only use the thumbs up when I get a really good result.
As I have both GG and GPT, I'll also take my best so so image and run it thought the other program and usually get far better results.

[NW 2156]

Yah sometimes they do freaky things. Just tried to add weathering to my skin for a boxcar model of my own design. Got a picture of a real box car, not my skin edited. Tried rephrasing, adding that is it is a skin for a 3d model of a railroad boxcar, and still got the same photo back. Got one I'll post,

I call it a GenesisF7. Apparently this engine was in a collision and they gaffed an F7A unit nose to it.
Otherwise a pretty good photo like render.

I actually like the way it looks. Reminds me of the Australian Clyde/ EMD CL class locomotives. Or their Alco PA known as the 44 Class.

Robert

[ossie]

The next question (for me at least) is whether or not AI could, by learning how, for example, OpenRails .dds image files are constructed to "lay" on a model, then create a bitmap file of a model image file that would be convertible to a .dds or .ace file--effectively "painting" a model for OpenRails.

That's something I've already had a very short run at with predictably inconsistent/frustrating results. I think we've all accepted we need to do a lot of training with it to teach it exactly what we want. Showing it side, top and end photos of a hopper and explaining the relationship of those parts to the texture file got an unusable (out of scale, poorly placed) result but did show (I'm sure) that with repetition accuracy would improve. That's reskins.... but for painters it's probably already at a point where you could use it. Using the latest TS Railbox pack I'm sure if you gave it a texture file and said "I'd like the reporting mark number changed from #### to ####, some different weathering marks added, some light rust patches around the door hinges, a large scratch in the paint across the Railbox logo..." etc you could probably get a new boxcar out of it very quickly.

[ossie]

I thought I'd test my theory about editing painted textures. I don't have the new TS set yet so I worked on one of the existing RBOX wagons (Shasta I think?). It took my daily quota via chrome then firefox to get it to this point. I ran through my chrome freebies just trying to convince it not to add to or change the layout of the texture file (1st result attached) and the last 2 attempts on Firefox have it pretty well spot on. I initially showed it a photo to use as a reference which it kept trying to do a sort of "reskin" with, changing the layout, even adding wheels at one point. Once I stopped referring to the photo reference and just zeroed in on instructing it to "make no changes to norm.bmp other than give it the appearance of aged and weathered metal" progress was made. I'd definitely still want to tweak it a bit before using it, but having now got the AI convinced to only work on what is there, not add, resize etc, it will be interesting if I try the same thing tomorrow with a different texture to see if it picks up from where we are now or whether it wants to go back to changing the layout etc. Trial, error and patience.... AGAIN.


Gemini_Generated_Image_66v6p966v6p966v6.png
norm.png

[yellowdog]

Just a reminder that uploading copyrighted material into an AI model may be a violation of the copyright.