AI-generated (not entirely realistic) railway images

[eldomtom2]

My general experience is that AI generation of train-related images falls down heavily if it's asked to produce a side-on view or anything outside of a three-quarter view. Often it'll just not do it and generate a three-quarter view instead.
This is an example of the flaws of how AI image generation works - most people with basic drawing ability, given a single three-quarter view of a train, could produce a decently accurate drawing of that train from a side-on view, allowing for some detail and proportion errors. But AI can't do that, because it needs a lot of images of something to be able to consistently reproduce it, and three-quarter train photos massively outnumber other train photos.

 

[Sun Chariot]

I asked it for a Freightliner Class 90, and I got...
Image shows AI blue class 66

Gemini AI clearly needs a loco recognition book!
I described a BR blue class 56 locomotive hauling a rake of HAA coal hoppers. Then, a Class 45 Peak 45106 hauling a passenger train through a busy railway station. Finally, a Gresley A4 passing over The Forth railway bridge.

And - well

 

[Brubulus]

I asked it for a Freightliner Class 90, and I got, well, something similar to what I usually get when I want to see an actual Freightliner Class 90 on a container train:

Image shows AI blue class 66

My favourite part of that is that it stuck Class 90 on the front of the 66.
It's like sticking "Class 390" on the front of the pretendolino.

 

[Class15]

My favourite part of that is that it stuck Class 90 on the front of the 66.
It's like sticking "Class 390" on the front of the pretendolino.

Haha I’m sure it can be made! That would be hilarious!

 

[contrex]

...that Thalys TGV seems to be formed entirely of power cars!

I bet it goes like something off a shovel, provided the feeder stations can keep up.

 

[Sun Chariot]

I bet it goes like something off a shovel, provided the feeder stations can keep up.

Let's hope they get those pantographs actually touching the wires too!

These interpretations of a Black 5 aren't too shabby.

 

[dosxuk]

The fun part is that AI training materials are already being infected by the AI generated output, so we're probably nearly at, if not already there, the peak accuracy of LLM AI generation. The more of these "looks like a train, but really isn't" images get posted, the more get read in as sample images of what a train looks like, meaning the inaccuracies are more likely to occur rather than less as we might assume.

It's the same with generated text - as more and more generated text gets read in, the more the generated output will look like it's been AI generated. Not good for those who are already coming to rely on AI tools...

 

[Sun Chariot]

My general experience is that AI generation of train-related images falls down heavily if it's asked to produce a side-on view ... because it needs a lot of images of something to be able to consistently reproduce it, and three-quarter train photos massively outnumber other train photos.

That's something I hadn't thought about and yet it's blimmin' obvious when I think about it! You're absolutely right and I have just spent ten minutes trying to generate a side on view of a Stanier Black 5. Not one single AI image; in fact, the only AI image I received, that wasn't a front 3/4 view, was this rear 3/4 view.

 

[johnnychips]

A class 333 on Dinting viaduct, courtesy of Deepai. It let me talk to the train, and it told me it really enjoyed doing this route, using electricity from the overhead wires.

 

[Sun Chariot]

View attachment 183777
A class 333 on Dinting viaduct, courtesy of Deepai. It let me talk to the train, and it told me it really enjoyed doing this route, using electricity from the overhead wires.

I'd be worried about the structural integrity of that viaduct, too...

 

[Russel]

I hope you're all proud of the fact that you're adding more "AI slop" to the internet, and in particular that some of these fanciful creations will probably appear as the results of image searches by people who would really prefer the genuine article!


(Yes, I realise that this is a tiny, tiny insignificant drip in the vast ocean of AI slop that's being added to the internet every day).

I'm hoping it's a passing fad and that people will get fed up of seeing AI generated garbage sooner or later.

The vast majority of AI images I've seen haven't been very convincing.

 

[Gloster]

I'm hoping it's a passing fad and that people will get fed up of seeing AI generated garbage sooner or later.

The vast majority of AI images I've seen haven't been very convincing.

…or they are sufficiently convincing that you don’t realise that they are AI unless something raises your suspicions and makes you check carefully. When it gets it right AI is dangerously good and it is continuing to improve itself.

 

[johnnychips]

Why are these results so poor? Considering AI can do me a scheme of lessons in literally three seconds, as per my post above #69, you would think it could do an image search for ‘Class 333’ and ‘Dinting Viaduct’ and put the two together. Of course, it can’t be that simple or it would have done it.

 

[dosxuk]

That's not how the current generation of AI works - it's just predictive text on steroids, rather than some tool going out researching what you mean and then producing a response - it's quickly thinking up what the most likely response to the question you asked is, and spitting that out. That's why you get suggestions like using glue to stick cheese to pizza, and why you get all these non-train trains.

It's a bit like asking a 5 year old to draw you a train. They'll get most of the bits in the right place, but it'll obviously not be a real train. The merry-go-round example above is a good demonstration of this.

It's always worth remembering that AI is as good at generating realistic trains as it is any other task you put it to - looks great at first glance, but failing in the detail.

 

[MarkyT]

…or they are sufficiently convincing that you don’t realise that they are AI unless something raises your suspicions and makes you check carefully. When it gets it right AI is dangerously good and it is continuing to improve itself.

It's a bit like bad photoshop illustration really, but at least someone was employed for that. Unless you're an expert in the subject matter, a vague likeness to 'a train' will often suffice for many commercial purposes. Members of this forum will comment on historical and logical discrepancies in the depiction of railway subjects in film and TV, and then there are all the continuity errors where sometimes characters seem to teleport between different trains mid-journey! They wouldn't make such basic mistakes for car travel even if that model didn't come out in yellow until 1969! Many members of the general public with no more than a passing interest see steam and smoke, hear a whistle and the slam of doors and think 'yes that's a train' and they're quite satisfied.

 

[Sun Chariot]

To prove what several recent posters have explained -
These attached non-railway images made with Gemini Image Creation and a short description telling the software what to include.
I don't know why this subject's richer in its reference material than that for railway subjects; but I hope it demonstrates how effective current AI results can be.

 

[m79900]

I asked for an EMT 222 at Belper. The livery's alright, but it doesn't look as real as some on here.


== Doublepost prevention - post automatically merged: 13 Jul 2025 ==

And then a 507 in the early 80s, I thought this was pretty spot on to be fair.

 

[MarkyT]

And then a 507 in the early 80s, I thought this was pretty spot on to be fair.

Defiitely PEP family, maybe an especially high-density variant with 4 double doors per car side. They might have built these for Manchester's Picc-Vic project!

 

[contrex]

And then a 507 in the early 80s, I thought this was pretty spot on to be fair.

Not bad, but I can see a pantograph (on a DC unit?) a lack of third rail, only two cars, odd looking driver's door and side window, blue and grey reversed (livery), BR logo placement and size?

 

[MarkyT]

Not bad, but I can see a pantograph (on a DC unit?) a lack of third rail, only two cars, odd looking driver's door and side window, blue and grey reversed (livery), BR logo placement and size?

There were AC versions of the same basic PEP unit, and some dual-voltage ones. The tool replicated the distinctive body and cab shape detail very well, though struggled with side windows, doors and livery as you say. There are lots of PEP unit photos around online today as the remaining fleets are fast being withdrawn, so enthusiasts are photographing them while they can. If an AI is using basic shape cues to identify potential source material then it has likely matched a lot of these similar standard 3/4 images as being the same object.

 

[Gloster]

To prove what several recent posters have explained -
The attached non-railway image result is with Gemini Image Creation and a short description telling the software what to include.
I don't know why this subject's richer in its reference material than that for railway subjects; but I hope it demonstrates how effective current AI results can be.

I don’t know what instructions you gave the AI, but as the result looks like Buddhist monks fighting perhaps one could look to China. They are known to have long-standing differences of opinion with the Buddhists over, well, all sorts of things so there may be a lot of source material of, how shall I say, completely false origin that AI has drawn upon.

 

[Brubulus]

I asked for an EMT 222 at Belper. The livery's alright, but it doesn't look as real as some on here.
View attachment 183823

== Doublepost prevention - post automatically merged: 13 Jul 2025 ==

And then a 507 in the early 80s, I thought this was pretty spot on to be fair.

You can usually ask it to do a stylised image or a photorealistic one - shows why correct prompts are so important

 

[Zomboid]

AI is only as good as the question you ask it. If you wanted a 170 in NSE (for example) then you could probably help it by attaching a picture of a 170 and a picture of something in NSE.

 

[tfw756rider]

And then a 507 in the early 80s, I thought this was pretty spot on to be fair.

That doesn't show. It just has a tiny "image" icon and the text "raw".

 

[m79900]

That doesn't show. It just has a tiny "image" icon and the text "raw".

Odd that, it showed when I inserted it. You can see this through other user's comments on it.

 

[styles]

I asked for an EMT 222 at Belper. The livery's alright, but it doesn't look as real as some on here.
View attachment 183823

== Doublepost prevention - post automatically merged: 13 Jul 2025 ==

And then a 507 in the early 80s, I thought this was pretty spot on to be fair.

This will be sold at the next {insert name of over-hyped Derbyshire tourist village} arts and crafts market; on a coaster, poster, mug, and teaplate.

 

[Technologist]

The fun part is that AI training materials are already being infected by the AI generated output, so we're probably nearly at, if not already there, the peak accuracy of LLM AI generation. The more of these "looks like a train, but really isn't" images get posted, the more get read in as sample images of what a train looks like, meaning the inaccuracies are more likely to occur rather than less as we might assume.

It's the same with generated text - as more and more generated text gets read in, the more the generated output will look like it's been AI generated. Not good for those who are already coming to rely on AI tools...

Plenty of clever people are working on all these things, AI is the worst it’s ever going to be.

The problem of AI running out of images or being re-infected by previous AI generated text and images is pretty solvable.

Firstly the models are getting more and better deductive logic put into them both by researchers and by their user base.

Secondly we do most of the cutting edge of science and engineering on computers it’s not a massive step to let AI models train themselves on output from those tools. So if you only have training data from three quarter view images of trains let it play with trains in CGI/CAD software so it can extrapolate from the existing image dataset.

 

[Sun Chariot]

I never knew Network Southeast had a "low cost" livery but, yes, apparently Google Gemini AI knows more than I do...

Still, it knows when a pair of Tractors is the required motive power - but that wrong line working had better be known to Control.

 

[D6130]

I never knew Network Southeast had a "low cost" livery but, yes, apparently Google Gemini AI knows more than I do...

Still, it knows when a pair of Tractors is the required motive power - but that wrong line working had better be known to Control.

Looking at the rolling stock, they may be hauling a railtour in the Down direction on the Up and Down Main between Templecombe and Yeovil Junction.

 

[Sun Chariot]

I bet this driver was miffed, after climbing so deftly up the bodyside, only to realise the door wasn't where he was....

== Doublepost prevention - post automatically merged: 13 Jul 2025 ==

The previously unknown pair of "not class 47" diesels is seen hauling a charter train.
Window hanging has taken a sadistic turn: one man is visibly mooning his bottom out of the second carriage, whilst a poor young woman is dangled on the side of the leading coach.

== Doublepost prevention - post automatically merged: 13 Jul 2025 ==

It's amazing how "belching smoke" can get interpreted...

== Doublepost prevention - post automatically merged: 14 Jul 2025 ==

"Where's the fake diesel locomotive? Will someone please point it out."