AI for model generation

[Lawrence@NZFinescale]

This is a new thread pulled out of a digression elsewhere.  So some of this initial post is a repeat of that.

ChatGPT can be part of a workflow to 3D models but won't (currently) generate the model. Note that ChatGPT is quite capable of being untruthful (and frequently is). If challenged it will happily own up to it and, while this is amusing, it can be frustrating.

It's handy for generating the images that other AIs can use to produce models though.

Meshy, Hitem3D and Tripo (and others) can produce models.  Below is the ChatGPT image when asked for a 30s female passenger standing waiting for a train.  You can refine this as needed by asking for whatever suits.  Using this image in the other AIs produces the 3D models illustrated.  These are all high poly watertight printable models.  Not perfect by any means but perfectly acceptable in scales less that 1:64 and easily refinable in larger scales.

This was all done with free/demo versions and one could viably get to a printable figure in 5 minutes.

This is far from the last word on the subject.  Realistically I have only dipped a toe in the water and I anticipate that in coming months this early exercise will look inept. Notwithstanding that I've gone from a process that took days to one that takes minutes and virtually no skill. It has issues, particularly as it is not amenable to tight control, but is still extremely useful.

I've continued with printing the Hitem3D version and will report back.  I like the Meshy version more, but it is not available for free download.  I'm not averse to subscribing to one of these services, but there are a lot of them and deciding which to patronise is a challenge.

[Rail and Tie]

Timely discussion Lawrence!

I have just been investigating the exact same thing for making printable figures/ items from historic photos. I have only played a bit with Meshy's free version, but was considering a subscription if it works well. I had not heard about Hitem3D yet, so thank you for the introduction. Looking forward to your results from your first prints.

[Lawrence@NZFinescale]

Timely discussion Lawrence!

I have just been investigating the exact same thing for making printable figures/ items from historic photos. I have only played a bit with Meshy's free version, but was considering a subscription if it works well. I had not heard about Hitem3D yet, so thank you for the introduction. Looking forward to your results from your first prints.

Thanks - I assume others are looking. The principle is dead easy, but I'm interested in anything that improves the process, makes it more controllable or speeds it up.

Generic people are useful, and fairly easy I think.  What makes custom figures more fun is the ability to create custom cameos, which I suspect will be trickier.

[Lawrence@NZFinescale]

So here she is painted.

She doesn't stand cruel enlargement but at just over an inch high she's amazing.  I'm fairly new to figure painting and while I think I have a grasp of the basic techniques my execution can be wayward. My first attempt was a bit rough, so this is a partial rework.

The facial detail is there.  It's beyond me to paint as such, but some washes have highlighted the detail enough to give a good impression.

There's a few hours in this paint job and, while I think I could do better, I don't actually need this figure so I'm going to leave her as she is and move on to some stuff I do need.

[Rail and Tie]

That turned out very well.  The figure detail is spot on, especially for the scale. I have been trying to find a way with the AI stuff to exaggerate the creases/folds in fabric which can help painting in small scales. No luck yet, but will be trying some of this this weekend if I get some time.

[Lawrence@NZFinescale]

That turned out very well.  The figure detail is spot on, especially for the scale. I have been trying to find a way with the AI stuff to exaggerate the creases/folds in fabric which can help painting in small scales. No luck yet, but will be trying some of this this weekend if I get some time.

I was pleased!

This AI route seems to be quite good at such detail.  However, painting in detail works well as well.  The crease/seam on the back of her jacket is painted in, not modelled.  This is the established idea of using a base tone, base+black and base+white. The idea is to paint in creases etc using the lighter/darker shades of the base.  It works very well.  It helps if the detail is there, but where it isn't you can easily add some.

I've moved on a bit.  I purchased subscriptions to Meshy and ChatGPT and I'm developing a workflow to produce specific figures that I'll write up at some point. What I want to achieve (and I'm basically there) is to start with an image (generally a period pic in my case) and to reproduce that figure convincingly in my desired pose without a lot of work. It's pretty simple and quick, but the AI steps take time (a minute or two each iteration), and it can take a few goes to get something that I'm happy with.  Even so, it's realistic to be ready to print within 30 minutes.

[Lawrence@NZFinescale]

I have been trying to find a way with the AI stuff to exaggerate the creases/folds in fabric

It may be that one AI will do this better for you than others.  Hitem3D seems to generate sharp features that catch paint more than something like Meshy which is generally smoother and produces nicer details (in my opinion). If you are working in HO then detail levels are probably moot as all the AIs will give you more than you can usefully print/paint.

[finescalerr]

Thank you for what is becoming a very useful clinic on AI for modeling. -- Russ

[Hauk]

A very interesting thread!

I think a hybrid approch would be the way to go for me. I need figures for my workshop diorama. I have some period images that I can use as a starting point. This fella has the right outfit, but the pose is a bit passive:



Then I used ChatGPT to get him working on the lathe:



Clothing and details seems to have preserved pretty well. Regarding the lathe, I feel that ChatGPT is not yet up to the task...

I played a little with my Hitem3D free credits during my lunch, break and things indeed look promseing:



All in all this little excercise took under 30minutes, including writing this post. With a little tweaking both of the images and the 3D model, this could really be something.

But really bad news for companies for Modelu...

[Hauk]

I ran out of files, but as you can see, the Overalls of the worker has turned into a jacket during the prosess. Here is a screen dump of the .stl:



So I went back to ChatGPT to make a better starting image. I will look more into this workflow, but It seems like the best route is to ma a set of images in ChatGPT that shows the figure from more sides. For now, I had only time to make the overalls more detailed:



What a weird time to be a modeller...

[Lawrence@NZFinescale]

Hi Hauk, you've latched on to my latest approach.

I've subscribed to ChatGPT and Meshy, which does open up possibilities and speed things up quite a lot.

My broad approach now for a specific figure is as follows (and close to yours) - but it does vary a bit case by case. For generic figures it can be better to give the software more free rein as you get a smoother result with less artifacts.

• Find a reference image
• Use ChatGPT to isolate the figure, repose and generate multiple views.  Multiple views may not always be needed, but where the unseen side is not obvious from the original image, additional images may provide the data.
• Use the references in Meshy. Try a single view first, or the multi (paid option) view if needed
• Meshy gives you 4 free retries.  This is useful as they differ a bit and one is generally superior to the others for my purposes
• Export stl without resize (important to preserve poly count)
• Tweak in Blender if required.  Typically adding a bit of additional depth to feet is useful and some simple sculpting to touch up details if needed.

[Lawrence@NZFinescale]

Hi Hauk

Based on your last reference image, Meshy delivered the attached. It took longer to sort the images for this post and write it than it did to produce the models! Each model is 17-18MB, and nearly 400,000 polygons, so plenty of data. At 1:64 they would be great.  I think they would be good in larger scales too. More detailed than many offerings, but some of the artefacts might be more obvious.

The first view is in my slicing software, which doesn't always render smoothly.  The other two images are the same figures in another viewer.

As you can see each try differs a bit.  The software seems to struggle with headgear and the caps could do with some some additional sculpting.

Experience suggests that these would print well.  They generally look better too as these viewers tend to distort the perspective a fair bit.

I'm happy to share the models if they are useful to you.

[Rail and Tie]

Awesome work Hauk and Lawrence! You have set me up nicely for success. It seems like this is taking off in the MRR world as well. I see that Doug over at Foscale models has released his first figures. I suspect using this exact same method in his "Humanity to Scale" series.

https://fosscalemodels.com/collections/humanity-to-scale

Also Bernard over at miniprints offers a 3D figure from a photo which I am sure uses meshy.

Things are moving pretty fast in this AI space.

[finescalerr]

I'd better learn to deal with AI because, in a couple of years, it may be possible to design a complete locomotive (kit) and all kinds of other goodies. I suspect, though, the whole process (from photos and/or drawings to AI to 3-D printing to cleanup and assembly) may result in some of us doing computer work and others focusing on assembly and finish. However things progress, it's inspirational. -- Russ

[Lawrence@NZFinescale]

I suspect, though, the whole process (from photos and/or drawings to AI to 3-D printing to cleanup and assembly) may result in some of us doing computer work and others focusing on assembly and finish. However things progress, it's inspirational. -- Russ

That ship has sailed.  The great thing about AI (for this application at least) is that there is MUCH LESS computer time required.

For things like figures where dimensional accuracy is not tremendously important it's really useful.