A practical use of Gemini AI for Modeling

[scottb613]

Hi Folks,

Just a practical use of AI for a simple model - that would be an easy first project for a Blender model.

Make me a low poly image of a typical bus - long haul - from 1950.

Bus1.png

Nice - make me a reference scale drawing of this - front view, side view, back view, top view - suitable as a reference in blender.

Bus2.png

Can you take the previous scale drawing and apply photorealistic greyhound textures to all four sides - so I can apply it to the model - as textures - in blender?

Bus3.png

Everything you need to build a simple low poly - nicely textured model - from 5 minutes with Gemini.

Regards,
Scott

[jbtower]

Amazing

[wacampbell]

This is an excellent way for us to leverage AI to improve our virtual worlds. I have done something similar, passing Gemini an old black and white photo of a boxcar, and asking for a 3d texture sheet. Its almost magical what it can do.

inGemini.jpg

[scottb613]

Hi Wayne,

Nice - yeah - anything that will speed model creation will vastly benefit all.

This is likely to put my old way of weathering cars out of business - I would have sometimes 25 layers of weathering in PSP - that would take me countless hours to produce - for a single freight car. Most of my models are all hand painted and weathered.

A couple things I’ve noticed - is AI is poor at recognizing numbers and lettering - I think it just treats it as part of an image rather than true character recognition - I’m sure it will only get better.

I haven’t checked actual dimensions - so I’m not sure how accurate AI is with that - which is why of been experimenting with non-critical items - vs rolling stock.

Now if we can only get more people involved.


Regards,
Scott

[eric]

Botched letters agr t#@ telltlea sgnn of AI imgray

[CSXT Q-317]

Awesome, I have found AI has some benefit in adding weathering effects on textures, and can even change the number on lets say a locomotive. I tried the below as a test. Not a bad first step.

Screenshot 2025-12-29 125906.png

[NW 2156]

I gotta admit, I've always been skeptical of AI. As this technology can be used for both good and bad. But I definitely can see how using AI to our advantage will help to only improve our hobby. And for that I'm grateful we live in the times we do.

Robert

[pwillard]

I tried the same bus prompts with ChatGPT. They were not as good as yours.

So I tried the same prompt with Gemini. and the results were worse than ChatGPT... so "it depends" and YMWV.

[scottb613]

I tried the same bus prompts with ChatGPT. They were not as good as yours.

So I tried the same prompt with Gemini. and the results were worse than ChatGPT... so "it depends" and YMWV.

Hi Pete,

Perhaps we are training AI in our individual sessions - I just took the same prompts I used above - changed it to a 1940’s ford pickup in a new chat window - and this is the result. Granted - the shadow is less than perfect.

IMG_3317.png


Regards,
Scott

[JSMR]

Whats the trick? I did the exact same phrase as the OP and got this in Gemini :

Gemini_Generated_Image_u0f506u0f506u0f5.png

Then the next one I did asking for photorealistic and got the exact same as this anyway and not a photorealistic image. lol
Gemini_Generated_Image_2ry8md2ry8md2ry8.png

[wwhall]

^I suspect that Gemini has not "learned" enough about you to understand your preferences. It does not "remember" images, it remembers what you have previously asked it to do. You might try this: Say something like, "I'm imagining a photorealistic scene of a 1950's-era U.S. Greyhound Scenic Cruiser bus built by General Motors sitting in unobstructed view of the photographer in a paved parking lot . Can you render that?" Then, in the next prompt, say "Can you now make me a reference scale drawing of this bus - front view, side view, back view, top view - suitable as a reference in Blender?" I just tried that and I got what I asked for.

[JSMR]

ok makes sense. I have barely used it. That's the 2nd image above. The 1st looked like a kids cartoon bus from 20 years ago lol. So some improvement.

[EldoradoRailroad]

While I find this quite "fascinating" it is not clear:
A) if a model was produced that can be imported into Blender?
B) if the textures "generated" that look "nice", but what do they look like on the model in OpenRails?
C) there are some posts here that point to non "AI" generated textures being weathered, but what does it look like on the model, in OpenRails?

I worry that this thread is going to be hijacked into "AI" generated images that are basically in the same vein as another thread in this forum, please don't! It would be quite sad if the images posted will just be "AI" generated "perfections" that have nothing to do with OpenRails.

Whenever I look at the cost per token for "AI" I balk at the price, especially if you go for more tokens etc. Ouch! I hope that this kind of "AI" is used to enhance textures, but it still remains to be seen if a detailed model can be created this way, and I mean a model that can be manipulated in Blender.

Happy 2026,

Steve

[pwillard]

The chat tools are best at "writing code" at the moment. They can do some texture art, but as you have seen, its kind of all over the map on the type of output it produces. Some might be useful, but you can just as easily get useless results... like front views labeled as side view images. They will tell you that they are not photo editors and prove to you that they are not very good at modeling. Its really too soon for it to just "make you a model of a GP40, with textures, and physics".

But at least you have a chance if you ask your ChatGPT AI to write a Blender Python script for you. Keep in mind, though, it really doesn't have much of a concept of making you a *good* model and certainly not a better one that you can make with a couple of hours of Blender training and practice.

[wwhall]

While I find this quite "fascinating" it is not clear:
A) if a model was produced that can be imported into Blender?
B) if the textures "generated" that look "nice", but what do they look like on the model in OpenRails?
C) there are some posts here that point to non "AI" generated textures being weathered, but what does it look like on the model, in OpenRails?

I worry that this thread is going to be hijacked into "AI" generated images that are basically in the same vein as another thread in this forum, please don't! It would be quite sad if the images posted will just be "AI" generated "perfections" that have nothing to do with OpenRails.

Whenever I look at the cost per token for "AI" I balk at the price, especially if you go for more tokens etc. Ouch! I hope that this kind of "AI" is used to enhance textures, but it still remains to be seen if a detailed model can be created this way, and I mean a model that can be manipulated in Blender.

Happy 2026,

Steve

I think that the unanswered questions that are being posed are these:

We know that AI can, by narrative, with and even sometimes without any reference image, create a "scene" that, for example, utilizes a certain model of locomotive. I believe that Scott has also already demonstrated that AI can take a Blender image, paint it in a certain livery, and place it in photorealistic image of a place or time. I think the question that is now being explored is whether or not AI can take an image--its own created image or another one--and turn that into output that could be used to create an accurate model in Blender. 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. I'm not a model builder, nor have I constructed .ace or .dds files from scratch, so I don't know the inner workings of that process for modelers. Call me an ignoramus on that aspect of modeling.