Railfan Technology > AI Tech vs K64 scan

I decided to scan an overexposed K64 of 5 Conrail SD45-2s taken at Cresson in 1990 and then ran it through Copilot with a simple command of "fix image". It did some decent cleaning up of the colors but little details like the number board font being changed and the colors reversed, partial removal of the cab-mounted Leslie RS3L and changing the radiators make it laughable and unusable. I assume I could try more specific commands or wait until the technology improves.

The third photo is another attempt where I said "fix sky only". It kept the numberboards, the horn and the correct radiators but derailed the first motor, and added funky trucks to the SD40 in the background and screwed up other things the first AI edit didn't.

Either way it raises questions of enhancing photo ethics and how much control AI has over our images.

Don Kalkman



Edited 2 time(s). Last edit at 04/11/26 05:00 by Mastadon.

---

---

---

Over the past several months, I've experimented with AI image generation, most all of it in "non-rail" subjects, but I've learned a huge amount of how it works and how it behaves.  Most of my experience has been with Google Gemini, but other AI platforms seem to act similarly.

• AI has only limited knowledge about what it "sees".  Not surprisingly, the more common or standardized an object is in the photographic or visual arts online ethos  (a commonly photographed automobile model, for example), the more accurately it can render it.  The more uncommon or unstandardized that an object is, the more issues that it will have rendering it accurately.
• AI has a nasty habit of "hallucinating" what it can't figure out when it renders an image and it won't tell you when it does it.  It can have particular problems with non-standard print fonts, trademark graphics, etc. if it does not have a clear image provided to it that shows the graphics, etc. in high detail.
• When AI does not correctly render details in an image and the user "prompts" it to correct the details, the AI may or may not correct the issue, but it may "bork" formerly other clear details in the subsequent render.  This is akin to making a photocopy of something, then running the photocopy through the copier again as the original.  With every copy (rendering), more detail is lost.
• If the user submits an image for the AI to use as a guide to create a render--for example, submitting a "roster shot" of a locomotive, then prompting AI to place the locomotive in a scene--AI may or may not do so successfully. AI often lacks "logic."  For example, prompting AI to place a steam locomotive with its tender spotted at a water tank for filling with water may result in the locomotive being completely misplaced, or out of scale with the water tank, or other issues.
• Many AI programs have built-in "blocks" to prevent it from being used as simply a photo-editing program.  For example, writing a prompt that says, "Change nothing in this image except to make the sky blue instead of gray", may generate an error where AI states, "I'm sorry.  I'm not a photo-editing platform."  Writing the prompt to say, "While leaving most of the scene intact, imagine the scene with a clear blue sky rather than a hazy, gray sky." Now you have asked AI to "think" about the scene and to implement what it believes is a change in weather conditions, not just change a color in a photograph.
• Even when you prompt AI to leave large element of a scene--the object being rendered (say, a locomotive based on an image that you provided it)--intact, not modifying perspective, etc. AI may still modify the perspective, orientation, etc. in the image it renders.  I believe that AI does this, in part, to avoid appearing to reproduce a proprietary or copyrighted image.  AI may also refuse to reproduce certain trademarks, etc. unmodified, or--in some cases--refuse to produce an image at all.
• One interesting feature of AI--if it refuses to do what you want, ask it why it refused.  Quite often, it will often give you the answer, and sometime even suggest a workaround.
• Seldom will AI accurately reproduce a specific location.  For example, I asked AI to render an image of a specific location in the downtown of a large, well-known city.  It rendered an image showing most every major building in that specific city location, but none of them in their actual location related to each other.  It treated the buildings like pieces of a puzzle that it did not know how to put together, so it just "hallucinated" a scene.  The scene looked nice, but was total fiction compared to the prototype location.
• Do not expect AI to understand what you understand.  For example, I prompted AI to render an image of a truck stop.  The rendered image looked nice--the buildings, fuel pumps, etc.  But it rendered the pumps in a way that no semi-truck could possibly pull up to them.  The truck stop had no entrances from or exits to the highway or streets around it.  AI understood what a truck stop is, it understands what its structures, etc. look like, but it had no "logic" or "understanding" of how or what the truck stop's function was or what its design was intended to accommodate.
• At a most basic level, AI has difficulty understanding direction--left, right, east, west, north, south, or sizes--meters, feet, inches, miles, etc.  For example, you prompt AI to render an image of a train traveling westbound on an east-west railroad at sunset.  AI sort of understands east and west as it relates to sun position, but it still may not get the image correct.  Now, ask AI in subsequent prompt, "Imagine the train now another 0.5 mile west  from its location and from the viewer in the last render."  All I can say is, "Good luck with that."  Your odds of getting that train another 1/2 mile west with that prompt is likely between slim and none.

I  can list more of my experience if its wanted, but this gives one some idea of the limitations of AI, at least the AI that is available at the "consumer" level.

In the example above, the AI-generated image has an odd number of rails!  And no its not an electrified third rail. 
As soon as I see a bogey like that, I lose all interest in the image. But I imagine it won't be long before the AI learns that each track has two rails.
 

hotrail Wrote:
-------------------------------------------------------
> In the example above, the AI-generated image has
> an odd number of rails!  And no its not an
> electrified third rail. 
> As soon as I see a bogey like that, I lose all
> interest in the image. But I imagine it won't be
> long before the AI learns that each track has two
> rails.
>  
Many "consumer level" AI programs don't really understand what they are "looking at"--the AI just "sees" pixels.  In the original photo above, the perspective did not clearly show the multiple tracks.  Your eye combined with your logic knew that there were multiple tracks there.  AI isn't that smart, at least not yet.  Here is a "fun" experiment to try:  upload a fairly basic photographic image to AI, then prompt it to imagine some minor additions or changes.  After it renders the image, then ask AI to write a prompt for you to use if you wish to replicate the image.  Don't be surprised if it renders a prompt that can be a dozen or more paragraphs of prose.  That's a good way to "see" what AI actually "knows" and what it doesn't and just makes stuff up to render the image.  Don't be afraid to argue to with AI when it clearly gets something wrong and gets stubborn about it.  AI may actually realize that you are correct.  The potential downside of that?  You are "teaching" AI things that it didn't know.  For better or worse, that will make AI even more powerful going forward.  And NEVER blindly assume that AI is correct.  Quite often, it is not.  The really funny thing, AI understands slang.  If you tell AI that "you're full of crap" on something, it knows what you are saying.

I'd like to see what someone who is pretty good with Photoshop and/or ON1 could do, starting with the same original.  

This is after about 2 minutes of effort in PS - tough to increase the sharpness on a relatively low-resolution image (TO shots are max 1500 pixels across).

BTW what is that orange box-thingy by the fuel tank of 6665?



Edited 1 time(s). Last edit at 04/16/26 15:50 by rbenko.

---