Westlake Publishing Forums

These are impressive renders. I think the challenge comes when you apply paint to the print-outs.

Its worth checking out the work of hyper-realist artist Ron Mueck. In my mind the detail and emotion he incorparates in his figures/sculptures is the equivalent of what we are seeing here with machinery and landscapes.

When I had ChatGPT generate more text-to-image content the next day, after a few minutes I received a notification that I had reached my daily limit and wouldn't be able to generate any more images in the free version for another 24 hours...
The same thing happened the next day... but this time I received an offer for a free one-month trial with all the features of the Pro version unlocked, which I could cancel at any time! Otherwise, a monthly fee of €23.00 applies.
That sounded tempting, because another 24-hour wait is quite a long time ;-)

For other projects, such as a harbor crane, fire station, Kemna Gigant roller, ... I'd really like to have a few custom figures that stand out from the little Preiser figures!

So I started with firemen (hose crew, attack crew, driver, engineer, fire chief, ...)

The initial request was:

Create a high-resolution file of a German firefighter from around 1955 wearing a classic black steel fire helmet with a leather neck guard.
The file is to be printed at a scale of 1:22.5 using a resin printer.
The original figure with the helmet is 180 cm tall.


After several changes to the request, I received the hose crew:

Feuerwehr_Mann_001 (fspg2)

generated with ChatGPT

Here, too, it is striking that the three men were given the same face.


Then the firefighter was depicted alone with the fire hose:

Feuerwehr_Mann_002 (fspg2)

generated with ChatGPT


The next command then generated the figure without a hose, with a similar hand position, so that the printed figure could later be fitted with its own hose and a nozzle.

Feuerwehr_Mann_003 (fspg2)

generated with ChatGPT


After that, I asked about the "best" slicers for resin printing:
https://3druck.com/programme/slicer-resin-harz-3d-drucker-28101482/

or an English-language page:
https://all3dp.com/2/sla-slicer-resin-printer/


I experimented with the rather powerful Chitubox and was amazed at how well the new feature, ChituGlyph, generated a 3D object from a JPG file.
 
Unlike TRIPO3D, it was virtually flawless. However, I could only create two files without incurring additional monthly costs (7.9 § / 8.9 €).

The alignment and automatic support structure generation can still be significantly improved manually!

Feuerwehr_Mann_004 (fspg2)

Figure created with ChituGlyph




All in all, you have to decide whether the monthly cost of around €20.00 for the full versions of Chitubox or LycheeSlicer is worth it!!!


However, there is also a very powerful and completely free slicer: ElegooSatelLite
https://www.elegoo.com/pages/satellite-3d-slicer

Here's a video where you can switch the audio track to different languages:
https://www.youtube.com/watch?v=nXBL2o3pODw

My first experiments were almost self-explanatory!
The automatic support structure was adjusted in no time:


Feuerwehr_Mann_005 (fspg2)

edited with ELEGOO SatelLite


Feuerwehr_Mann_006 (fspg2)


edited with ELEGOO SatelLite

Elegoo makes some very good printers. Offering a non-proprietary "very powerful and completely free slicer" seems generous and a clever way to promote their printers, a very rare approach. -- Russ

Kinda, sorta, a bit like Blackmagic Design and the free version of DaVinci Resolve. The software helps sell their hardware.

@Kevin
That's a fantastic program. For the last six years before I retired, I had the opportunity to support the sales of Blackmagic Design products as a channel manager.

Over the past few days, I've had quite a few surprises while creating various 3D figures using ChatGPT.
Depending on how I phrased the prompt, the results were sometimes very strange.

For example, after making several changes to the prompt, I received this image:

Feuerwehr_Mann mit Krawatte (fspg2)



However, I didn't want a tie, so here's the new version:

Quotewithout a tie

Then ChatGPT sent the following message again:

We're sorry, but the image we generated may violate our safeguards regarding harassment, discrimination, bullying, or similar prohibited content. If you think we made a mistake, please try again or edit your prompt.

My question:

QuoteWhat is the correct phrasing? You're not making this easy for me.

Yes—the problem is, unfortunately, that even small terms like "German steel helmet," "command stance," "deployment," or even "remove tie" sometimes trigger a false security filter.
Very neutral model-making descriptions work most reliably.

This description should work much more reliably:

Historical civilian firefighter model figure in a dark fabric uniform without a tie, traditional black firefighter protective hood with neck protection, dynamic working pose with open hands for separate accessories, highly detailed resin model figure, 1:22.5 scale, isolated against a neutral white studio background, realistic fabric folds, stable posture for 3D resin printing, documentary model-making style, no military depiction.

Important changes:
•    Replace "helmet" with "firefighter's protective hood"
•    Avoid using "steel helmet" entirely
•    "no tie" directly under "uniform"
•    "open hands for separate accessories" instead of hose/nozzle
•    "no military depiction" explicitly added
This significantly reduces false triggers.

So I adopted the suggested prompt:

QuoteHistorical civilian firefighter model figure in a dark cloth uniform without a tie , traditional black firefighter protective hood with neck protection, dynamic working pose with open hands for separate accessories, highly detailed resin model figure, 1:22.5 scale, isolated against a neutral white studio background, realistic fabric folds, stable posture for 3D print, documentary model-building style, no military depiction.

... and finally received this image:

Feuerwehr_Mann ohne Krawatte (fspg2)


However, ChatGPT labeled the image "Soldier with helmet"... apparently, it's not military


Over the past few days, I've not only been able to create images of several figures, but I also had Kalle print some 3D STL files generated by TRIPO3D on his Formlabs 4.

That's how I ended up with a 78mm-tall printed figure based on the STL file I originally created of my father:

ChatGPT_Test_09 (fspg2)



ChatGPT_Test_10 (fspg2)




The first firefighter was there too:

Feuerwehr_Mann_007 (fspg2)


Feuerwehr_Mann_008 (fspg2)



The roofer also tried laying the first few roof tiles as a test:

Dachdecker_06 (fspg2)



This finally allows me to create the desired shapes using relatively simple methods.

Dachdecker_07 (fspg2)


I haven't sanded any of the printed figures yet—you can hardly see any traces of the support structures!
Only when you zoom in (click on the image—click again in the gallery that appears to see the 100% view).

Compared to some figures I had printed back in 2017, there's a huge difference in how "invisible" the print lines are!

That roofer is something else. The print and model are good, but it's the way it enables you to tell the story that makes him fantastic.

Those are some terrific figures.
It looks like you are quickly learning how to give AI effective and "acceptable' prompts to get amazing results.

OUTSTANDING! And I strongly agree with Lawrence about the roofer.

Just as you, I've run into AI having some odd perceptions. One time I tried to clarify something and AI told me to contact a suicide prevention service. I had no idea AI had such strong feelings for me ....

Russ

Until now, I have viewed artificial intelligence with a certain degree of ambivalence.
On the one hand, I clearly recognize the impact it will have on numerous professions and fields of work. On the other hand, I am concerned by how readily AI-generated content is often accepted as "true" or reliable, even though it frequently contains errors or presents information in a convincing yet ultimately incorrect manner.

At the same time, over the past few weeks I have become increasingly impressed by the possibilities this technology offers.
When I think back to the many hours—sometimes even days—that I spent as a photographer creating complex image composites, much of what is possible today seems almost unbelievable. With precisely formulated prompts, images can now be generated within minutes that would previously have required a considerable investment of time and effort.
The results often impress me with their creativity and technical quality. At the same time, however, they challenge me to look more carefully and critically evaluate what is being presented. The more convincing the results become, the more important it is to distinguish between an impressive representation and actual reality.
It is precisely this tension between fascination and critical reflection that makes the subject of AI so interesting to me at the moment.

For example, since 2018 I had been searching for information about the lettering on the old Münzel crane. After all, I did not want my scale reconstruction to carry an imaginary or inaccurate inscription.
In the enlarged detail shown below, taken from an old photograph belonging to my father, deciphering the lettering had never been satisfactorily possible with any magnification tool available to me. Various image-enhancement and sharpening tools also revealed only indistinct and barely recognizable characters.


Münzel Kran-Beschriftung_01 (fspg2)



After I had already contacted various archives, museums, private individuals, and even the local newspaper in search of further information, I was fortunate enough to receive two photographs from the Elbe Shipping Archive in Lauenburg (today the German Inland Shipping Archive). These images were primarily used for the reconstruction of the crane and the sand bunker that once stood beside it.
However, the exact lettering remained a mystery until quite recently. The only inscription I had been able to decipher was "Außenhafen" ("Outer Harbor") on the rear side of the crane house.
All the more delighted I was when ChatGPT pointed me toward the Facebook page of the Heimatbund Lauenburg. There I actually discovered two wonderful photographs of this crane, both showing the lettering clearly and legibly.

The first photograph shows the original version of the boom-luffing mechanism, featuring the large cable pulley used for adjusting the boom height. It also reveals the old company address, which can still be read clearly:


Walter Münzel
Hafen - Umschlags – Betrieb
  Reederei – Bunkerstation
        Lauenburg/E
Hafenstr 15.  Telefon 280

Münzel Kran-Beschriftung_06 (fspg2)

https://www.facebook.com/HeimatbundLauenburg/photos?locale=de_DE


The second photograph shows the later version of the crane, equipped with an electric drive for adjusting the boom position.
The lettering had also changed by this time.
This image documents a later stage in the crane's history and provides valuable evidence of the modifications that were made over the years, both technically and in terms of its external appearance.

Walter Münzel
Hafen - Umschlags – Betrieb
       Bunkerstation
Kies u. Mörtelvertrieb

Münzel Kran-Beschriftung_02 (fspg2)

gefunden auf: https://www.facebook.com/HeimatbundLauenburg/photos?locale=de_DE

In the 100% view of this photograph, the halftone pattern from the newspaper printing process becomes clearly visible.
A further note regarding the enlarged images: Clicking on the respective photograph opens it in the gallery, and a second click displays the image at full 100% resolution.

Driven by curiosity, I then asked ChatGPT to remove the newspaper halftone pattern from this image and to apply a careful colorization:

Münzel Kran-Beschriftung_03 (fspg2)

gefunden auf: https://www.facebook.com/HeimatbundLauenburg/photos?locale=de_DE

The colorized result is quite convincing at first glance and creates a remarkably realistic impression. But what had happened to the lettering? 

While the overall image quality appeared to have improved significantly, the inscription on the crane had changed in a way that immediately raised doubts. This was a striking reminder that AI-based image enhancement and restoration tools do not merely reveal existing information—they can also interpret, reconstruct, or even invent details when the original data is unclear.

For me, this was a perfect example of both the strengths and the limitations of artificial intelligence. The image looked more realistic and visually appealing, yet the very detail I was most interested in—the historically accurate lettering—had become less trustworthy. It demonstrated how important it is to verify AI-generated results against original sources whenever historical accuracy is required.

Münzel Kran-Beschriftung_04 (fspg2)

gefunden auf: https://www.facebook.com/HeimatbundLauenburg/photos?locale=de_DE

The subsequent AI sharpening altered further details as well. For example, the number plate of the truck suddenly displayed different digits.

So I asked the AI bot how I should have worded my prompt in order to ensure that everything appeared correctly and remained unchanged.

The answer was very detailed:

"Historical restoration of a black-and-white photograph"
Preserve the original subject matter absolutely unchanged.
Perform only the following operations:
• Remove halftone screening and moiré patterns (descreening).
• Carefully reduce film grain.
• Optimize local contrast.
• Brighten shadows and dark areas.
• Preserve highlights.
• Sharpen original contours.
• Improve the legibility of lettering, but do not add or invent any characters.
• Keep the original perspective unchanged.
• Do not reconstruct missing image information.
• Do not add new objects.
• Do not alter any people.
• Do not redraw faces.
• Do not add vehicle details.
• Do not modify the crane, cables, timber framework, windows, or lettering.
• Do not create artificial wood textures.
• Do not perform AI redrawing.
Then apply careful colorization:
• Use realistic color tones typical of the 1950s.
• Use natural, slightly muted colors.
• Colorize only where image information is present.
• Leave unclear areas neutral.
• No fantasy colors.
• No invented details.

Objective:
Archival-quality restoration and colorization while preserving every original detail.
Documentary reconstruction, not artistic interpretation.
Lettering, timber framework members, cable runs, bearings, window divisions, and sheet-metal edges must correspond pixel-for-pixel with the original photograph. Any geometric alteration is unacceptable.

This final paragraph is the key element that prevents many AI systems from modifying timber framing, crane components, or vehicle details. However, the fundamental problem still remains: generative image AIs cannot always comply with these instructions.
For truly accurate restorations of historical harbor photographs, traditional tools such as Photoshop (using Dust & Scratches, DeScreen, and High Pass techniques), Affinity Photo, or dedicated restoration software are often more reliable than generative AI systems.

Naturally, I wanted to put this advice to the test immediately. Using this prompt, I obtained the following image:


Münzel Kran-Beschriftung_05 (fspg2)

gefunden auf: https://www.facebook.com/HeimatbundLauenburg/photos?locale=de_DE

Here, too, there is a very serious error.

Münzel Kran-Beschriftung_08 (fspg2)



When I applied the same new prompt to this older photograph of the crane, new mistakes also appeared in the lettering.


Münzel Kran-Beschriftung_07 (fspg2)

https://www.facebook.com/HeimatbundLauenburg/photos?locale=de_DE

In summary, after generating a new image with an AI system, it is often worthwhile to ask the AI itself to evaluate the result for plausibility and accuracy.
Doing so frequently reveals additional weaknesses, inconsistencies, or potential errors, as well as suggestions for improved prompt formulations.
These refinements can then be incorporated into the next prompt and tested in a further iteration.

When creating 3D figures for resin printing and integrating the faces of real people, I have sometimes gone through five or six rounds of corrections and refinements before I was satisfied with the final result.

In that respect, AI is not so different from ordinary human beings—we all make mistakes.
The difference is that an AI will usually acknowledge those mistakes when they are pointed out and, in many cases, can even help identify and correct them.

Personally, I find this development fascinating. Even though the hours often seem to fly by while I work toward a satisfactory result, the process itself remains both interesting and rewarding. Each iteration brings new insights, and the combination of experimentation, critical evaluation, and gradual improvement is very much part of the experience.

Frithjof,
 Thank you for your detailed description of your experience working with AI photo restoration.
I have no experience with using AI for any modeling research or projects, but have seen how diligent I must be when evaluating any AI information.

AI seems capable of doing useful things IF great care is taken with the prompts it is given and the results it presents.

I have experienced the same issues as Frithjof. Over the past four days, for example, Google Gemini made so many errors I asked whether it could explain the reason. It was unable to do that. I suspect Google replaced its previous AI with a revision and the new version has some bugs. The current version actually apologized for its errors and thanked me for my patience. -- Russ

Quote from: fspg2 on Today at 12:49:20 AMIn summary, after generating a new image with an AI system, it is often worthwhile to ask the AI itself to evaluate the result for plausibility and accuracy.
Doing so frequently reveals additional weaknesses, inconsistencies, or potential errors, as well as suggestions for improved prompt formulations.
These refinements can then be incorporated into the next prompt and tested in a further iteration.

This is also my experience, and very useful it is - up to a point.

In life, it's often the case that when things seem stupid, it's that we lack insight as to the underlying systems or causes.  The AIs seem to be compartmentalised for various tasks and it is very helpful to ask questions about behaviour and how one might better approach the problem for a better result. They are 'self-aware' in that sense.

It's worth noting that AIs can often grasp the problem well, yet still be unable (for whatever reason) to deliver a useful result. They always deliver something though, even if 'I cannot do that' would be a better response. Sometimes it can be like dealing with an amiable idiot.

They also suffer from 'drift' (ChatGPT's term). They appear to weight a chat towards the most recent queries and can 'forget' earlier answers. They clearly carry SOME information between chats, but not all. So an amiable idiot with memory issues.

Probably we are dealing with the nature of the system. Traditional programming is rigid, but AI is fuzzy and flexible. For the things discussed here the AI is sitting a little too far on the fuzzy side. It is possible to constrain it with appropriate structures (You can ask it to generate a protocol and then save this for use in later sessions to get a more consistent response to similar queries), but possibly we are bumping into current limitations of the technology.