Westlake Publishing Forums

Embrace technology but don't abuse it. -- Russ

Your work deserves international exhibition. -- Russ

Whenever one of you people makes it to Germany's famous river Rhine don't miss to stop by at automata museum 'Mechanicum' in the even more famous town of Rüdesheim, where I'm showing some of my photographic work and a small selection of models (there are three altogether, so it's one or two, and one of those two is definitely 'quiet earth' ). The exhibition lasts from March 14 until September 30. Route description.









See you!

Volker

Quote from: finescalerr on March 01, 2026, 12:03:10 PMArticles I read suggest noticeable improvement may occur within one to five years. -- Russ

Undoubtedly, it's not hard to catch them lying.  As I noted elsewhere, the overall process was impossible 4-6 months ago, borderline 2 months ago, and stunning today.

It's possible to get very tied up in debates about whether AI is 'good', whether it is actually 'intelligent' and so on.  I prefer to ask 'Can this new tool improve my enjoyment of model-making?'.  Personally I feel that it has.   

I have tried "talking" to AI to better understand how it works. I've experimented with both Gemini and CoPilot and can confirm Lawrence's comments. Current versions of AI may retain all the data leading up to a given point but, when you point to an error, they only focus on the more recent details and may bluff their way through the rest. That would account for what Lawrence has experienced. Articles I read suggest noticeable improvement may occur within one to five years. -- Russ

Quote from: Stuart on February 28, 2026, 08:52:22 PMLawrence, yes I understand what you are saying. I too have been frustrated as I have tried to rework an image where parts of it were satisfactory and other parts not so. Try as I will, AI would not retain the portions I liked but would rework the image altogether usually with poorer results. I am not familiar with how to lock an image.

Stuart

Bearing in mind I don't really know what I'm talking about here, and I'm paraphrasing ChatGPT's responses when I've asked it similar questions:

The basic principle is to ask ChatGPT. After doing a few figures along a similar workflow I asked ChatGPT how I could formalise that workflow and it did that - Extremely well I thought, with a level of refinement and flexibility that I hadn't considered.  Then I made a few changes, introduced drift again and it all went pear-shaped.  But the principle worked very well.  The problem is that if you want to make a change to something 'locked' you need to be VERY specific. While this is a bit frustrating, ChatGPT will rework the whole thing in a few seconds if things go awry so it's not the calamity it would be in the real world (I haven't got back to my workflow yet, but I'm very relaxed I'll be able to quickly get to a result). So my advice is to ask the question above of ChatGPT and have it generate an environment that suits you.

I don't think its a fully resolvable problem though.  We tend to think of the process as a road.  If you take a wrong turn, you back up and take a different route.  However ChatGPT is more like a river.  You can back up, but meanwhile everything has changed a little.  As noted above it seems that you cannot return or lock an image as such - only to the AI's interpretation of that image.  Locking it avoids reinterpretation of the original image, but not reinterpretation of the AI summary.

A bit of discipline can help too I suspect.  I think it's a mistake to chat with digressions as one would in person.  If you want to digress, start a new 'chat' and come back to the original focussed thread when done.  I confess I haven't often followed this advice very well.

I'm quite familiar with this sort of problem as it's not unlike many spousal conversations that I have.  It's the crux of AI I suppose.  It needs to be flexible to be useful, but not so flexible as to be useless.

Lawrence, yes I understand what you are saying. I too have been frustrated as I have tried to rework an image where parts of it were satisfactory and other parts not so. Try as I will, AI would not retain the portions I liked but would rework the image altogether usually with poorer results. I am not familiar with how to lock an image.

Stuart

Quote from: Stuart on February 27, 2026, 02:53:11 PMUsing the same descriptive input to ChatGPT as the first image, this second image was created.

Stuart, your images show both the strengths and weaknesses of these tools.  The same input generates different output.  That's very handy when it comes to generating similar, but individual figures.  On the other hand if you are trying to refine the same figure it can be extremely frustrating. ChatGPT calls this 'drift' and it seems it doesn't use the image as such, but some sort of shorthand. As I understand it, from what the AI described, it analyses the image and prepares a 'text' description of what it sees and then works from that.  Every interpretation of the image might produce subtly different text and indeed every interpretation of the text produces a different result. Thus unless you some how lock and label a result, it's impossible to back up the bus if things go awry.

For instance, the AIs seem to favour US style headgear, which is inappropriate for much of the world historically.  If you get the perfect figure with inappropriate headgear it can be difficult to change the hat, without changing the whole figure a bit. Sometimes that really doesn't matter, but other times it sure does.

Clearly ChatGPT doesn't know the critical features of a spanner either :-)

Leonard Jones was a contributor to Finescale Railroader and worked for the Southern Pacific as a herder. He explains, "Herders move trains in and out of the yard, guide hostlers onto outbound trains so they're coupled up and air tested when the train and engine crews show up." But Leonard also builds working "models" in 1:1 scale. The photos show a partly scratchbuilt and carefully weathered (that's right, he weathered it) 1921 Harley board track racer. Leonard says, "Yes it runs and yes, I still ride it."

Maybe the photos will inspire somebody to build a scale model.

Russ

Quote from: Stuart on February 27, 2026, 02:40:35 PMHere's something I thought I would try.

...

This process might prove useful when a particular pose is wanted and is difficult to describe or for AI to understand fully.

Stuart

Brilliant!  Bloody Brilliant!!