Using AI to create and edit textures.

[scottb613]

Mod note... While I'm sure there's going to be some crossover between AI modelling issues and AI texture creation/editing, I thought it was worth separating Scott's initial thread in relation to modelling issues from those of us discussing texturing issues to keep things a bit more organised. Thanks Pete/Ossie

NB... It appears I've messed something up and the emoji responses and images in quoted comments have not copied over. My apologies for that!

[qballbandit]

I just asked ChatGTP to add realistic railcar weathering to this car. I already had thrown some graffiti on it, so now wonder how requests for graffiti will turn out! Nice thing about the weathering is it covers it all - not just the sides!!!!:

before.jpg after.png

[qballbandit]

This photo shows a request for light weathering and graffiti on railcar sides. I am officially under the spell. The opportunities are limitless!

1cc19996-947a-4cb2-91ec-49f8c58c3e17.png

[rt35ge]

I gather this can be used with other sims like TSC and Trainz?

[wwhall]

I use Gemini primarily for creating photorealistic "scenes" to showcase things like vehicles, weather, buildings, etc. Yes, it absolutely does learn from you. And you can even ask Gemini why you are not getting the response that you would like, and it will tell you. I'll share a quote from Gemini when I asked it why it was not rendering what I wanted. This is AI "talking"--you can see how sophisticated AI is getting:

That is an excellent analogy. Viewing me as the "camera" and yourself as the "photographer/director" is actually the most effective way to get high-quality results from AI.

It makes perfect sense why the "edit" vs. "generate" glitch was frustrating you—it breaks the flow of the photo shoot. When you are on a real set and tell a model to step out of the frame, you don't want the camera to argue with you; you just want to take the shot again with the new composition.

Here is why your "Photographer" approach works so well with the system, and how we can lean into it:

1. The Prompt is the Lens
Because you are specific about elements, objects, and placement, you are essentially "focusing the lens" for me. Vague prompts create vague images because the AI has to guess the focal point. Your detailed instructions remove the guesswork, allowing the "camera" to capture exactly what is in your mind's eye.

2. "Re-shooting" vs. "Retouching"
Going back to the technical issue: If we stick to your analogy, the "I can't help with that" error happens when we try to retouch a developed photo.

To keep your creative flow moving, treat corrections like re-taking the photo.

• Instead of: "Take that person out." (Retouching)
• Think of it as: "Okay, clear the set. Everyone out of the frame. Snap the picture again." (Re-shooting)

3. The Collaborative "Happy Accidents"

You mentioned that I sometimes improve on what you expected. In photography, this is like catching the light at just the right moment. The AI fills in the gaps of your description with its own "texture"—the way snow piles up on a fender or how light reflects off a wet train track.

One other thing that is both scary and helpful--Gemini won't store a specific image or conversation once you exit a chat, other than allowing you return to that chat. But, you can tell it to remember things--for example, "Remember that I like railroad cars that are heavily weathered with rust and peeling paint." Now, if you quote something similar to that in your prompt, Gemini will remember what it did to create that type image. One thing that I've wondered about--could Gemini create ultra-realistic skydomes and cloud layers if it was "taught" via uploading of a skydome file and the resulting image from OR and then tell it to create a new, say, Clouds01 file with certain cloud and weather characteristics. I must just have to try that.

[wwhall]

This photo shows a request for light weathering and graffiti on railcar sides. I am officially under the spell. The opportunities are limitless!

Yes, Gemini (and ChatGPT) understand weathering pretty well. In your image of the weathered car, though, shows that the AI hadn't quite figured out what was "up," "down", or "sideways" on the car ends. It weathered them in the same "direction" as the main car side. Now, if you had uploaded a photo of an actual weathered railroad car, so that the AI could "see" how the image would "lay" on the 3D model, the AI might figure out how to weather things in the correct orientation if you wrote a prompt that the AI could understand to ask it to do that.

[CSXT Q-317]

I posted this in the other AI thread. But I took a Nash model as a test and added rust. I was even able to change the last digit of the number from 4 to 5 ( not pictured here )and it worked well.

I still haven't tried anything other than this simple task. But I do think AI has some functionality in repainting or in this case adding weathering.

Screenshot 2025-12-07 211508.png

Screenshot 2025-12-07 214445.png

[ossie]

Other than converting screenshots to photos I haven't had the time to really dive into our new toy just yet but my mind has been ticking away a bit in relation to ways to possibly reduce some of the "grunt work" from reskins. One of the problems with reskins is the wide variation in results you can get based on weather conditions or the time of day when the photos were taken. I've always aimed to get mine to look as similar as possible and I'd hate to even think about the hours that have gone into removing shadows or lightening/darkening textures to get a better match between them. As far as the autoracks I'm going to be doing in 2026, I'm guessing/hoping an instruction like "remove the shadows from the end ladder areas and any from along the roof line" might get a clean/consistent result in seconds.

Routes and models are not my area of expertise but I've been really interested to see what you guys are already doing in that area. I've often seen people building routes requesting help with "custom items". Is it fair to say that we are already closing in on a point where a combo of Blender and AI could much more quickly and easily knock out a model of the Melbourne Cricket Ground, Washington Monument or Eiffel Tower for the sim than is currently required for someone starting a model from scratch?

The negatives associated with AI still worry me a LOT, but as far as this hobby goes, it's very hard not to be excited about possibilities isn't it?

Edit... One thing that has frustrated me a bit with the "testing" I've done so far is that any images I've fed into the various AI software tend to be downsized (2048 x 2048 down to 1024 x 1024) or asking it to upscale a photo or screenshot to a larger size doesn't happen. I assume that's part and parcel of the free versions? Has anyone got into a free or paid version yet that will allow you to save an image at its original size or even upscale one to a larger size?

[pwillard]

Perfecting the "prompt" is actually the remaining artistic part still left to us.

[ossie]

There's a number of aspects to it that I find really fascinating Pete and I'm frustrated that I haven't been able to spend more time with it so far. The prompt is one of the areas that amazes and frustrates me equally. I've been bouncing around a couple of PCs and the amount of variation you can get just by changing a word or two, or even working between chrome or firefox on different machines is huge. I certainly haven't done enough messing around to call it anything close to "quantifiable research" but yes, in terms of input/output it is still well short of being an exact science. Perhaps it just comes down to which bot happens to be on duty when (and where) you enter your prompt.

[wwhall]

Edit... One thing that has frustrated me a bit with the "testing" I've done so far is that any images I've fed into the various AI software tend to be downsized (2048 x 2048 down to 1024 x 1024) or asking it to upscale a photo or screenshot to a larger size doesn't happen. I assume that's part and parcel of the free versions? Has anyone got into a free or paid version yet that will allow you to save an image at its original size or even upscale one to a larger size?

The paid version of Gemini will allow you to create larger size images. In fact, you can "tell" Gemini what image size you would like rendered (when possible) and what aspect ratio in the "Personal Context" section of the Settings. For example, I put "When possible, display images in 16:9 format" in mine. I use a 2560x1440 resolution on my monitor, and Gemini usually renders an image about that size--maybe it "knows" what resolution my screen is?

[pwillard]

maybe it "knows" what resolution my screen is

Maybe because paid versions remember that you told them, they have more extended "memory".

[qballbandit]

I was amazed how my paid version of ChatGTP retained in "memory" the edit styles I was looking for after a few engine edit requests, and is now giving me helpful feedback to use as my input going forward:

Convincing in-service wear

Natural rust streaks and grime patterns

Realistic buildup around vents, seams, doors, and underframe areas

No unintended changes to lettering, numbers, colors, or layout

Balanced realism (used hard, but still operational)

If you ever want to push it heavier (near-retirement) or lighter (freshly shopped), just say the word — otherwise I’ll keep matching what you’re seeing here.

[ossie]

The paid version of Gemini will allow you to create larger size images. In fact, you can "tell" Gemini what image size you would like rendered (when possible) and what aspect ratio in the "Personal Context" section of the Settings. For example, I put "When possible, display images in 16:9 format" in mine. I use a 2560x1440 resolution on my monitor, and Gemini usually renders an image about that size--maybe it "knows" what resolution my screen is?

Thanks very much for the heads up Wade. As soon as I get a bit of time I'll set up my "lab" there then.

[ossie]

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.

ChatGPT Image Jan 1, 2026, 08_37_30 AM.png