3D print people figures

[Durley]

I just had a quick play using the same picture of SWMBO in a sparkly dress that I used in my original post on this thread (page 1). It seems to have done a good job with the detail, BUT, it has also added more than a few pounds to her... I ain't risking my life by posting it!

Wise move Phil!

I’ve tried using Google Gemini to create the 2D image from a photo and text description, which has worked better at giving the pose and attire I want. Then use Meshy to do the image to 3D model part.

Original photo of a GVT driver.

IMG_4896.png (1.72 MiB) Viewed 5879 times

Google Gemini cleaned up 2D version.

IMG_0129.png (557.01 KiB) Viewed 5879 times

Meshy 3D model.

IMG_0132.jpeg (126.49 KiB) Viewed 5879 times

I’m off to fire up the 3D printer to see how he comes out.

[philipy]

Cannot wait to see how this comes out! It's amazing so far.

[ge_rik]

Stunning!!
Can't wait to see your painted version.

Rik

[Sam95]

Wahouu that's amazing.

[Durley]

…and here he is fresh from the printer.

IMG_5331.jpeg (924.19 KiB) Viewed 5787 times

Wow was my immediate reaction

[philipy]

…and here he is fresh from the printer.

Wow was my immediate reaction

Seconded!

[ge_rik]

…and here he is fresh from the printer.

Wow was my immediate reaction

Seconded!

Thirded. I might try digging out some photos of Southwold Railway staff to see what happens.

Rik

[philipy]

I might try digging out some photos of Southwold Railway staff to see what happens.

I did try to have a play last night, but having to try to understand Gemini and how to use it to emulate what John did with his figure, defeated me, so I didn't get to the Meshy stage, let alone printing.

[Durley]

I’m not a Gemini expert but here’s what I have been doing:

1. Find a picture of a real figure, either take a photo from a page in a book or download from the internet.
2. Go to https://gemini.google.com/app
3. In the prompt box press the + button and upload your starting picture.
4. Open the settings menu by pressing the symbol next to the + and change to Create Image (image banana).
5. Type in a prompt along the lines of ‘Please (*) create a realistic image of this person. Remove the background.’
* ‘Please’ is optional but I was told on a recent AI prompting training course I did at work that research suggests polite prompts get better results. I don’t know if this is true but politeness costs nothing and when the machines take over the world they may look more kindly on those of us who were polite to them!!!
6. Once you have a Gemini created image, make additional prompts if needed e.g. ‘Make the hat look more like a traditional British working class flat cap’ etc.
7. Once you have an image you are happy with, upload to the 3D Model from image section on Meshy.ai.

[philipy]

Thanks.
I kind of understand the prompt writing, although it seems to need more written detail than I was expecting, given that it is working on a picture not creating from scratch, but useful to know about being polite to the machine!! . However I was more struggling with how you remove all the speckles, fuzz, etc.and improve the focus.

[Durley]

Thanks.
I kind of understand the prompt writing, although it seems to need more written detail than I was expecting, given that it is working on a picture not creating from scratch, but useful to know about being polite to the machine!! . However I was more struggling with how you remove all the speckles, fuzz, etc.and improve the focus.

Try a prompt something like ‘Create a high definition photo realistic image of this person please.’

I used that prompt plus ‘Please remove the background’ and ‘Show him standing up straight’ on this image of GVT driver James Hughes.
Starting image:

IMG_5335.jpeg (2.15 MiB) Viewed 5739 times

End image ready to upload to Meshy:

IMG_5340.png (552.51 KiB) Viewed 5739 times

You could prompt Gemini to show him with his arm out, holding something, sitting down, etc, etc.

[philipy]

Ta!
That looks/sounds fantastic. It is very impressive the way it has straightened him up, you'd never know from just looking at the last picture.

[philipy]

Just out of curiosity, I wondered what it would do and would it be able to "do a Meshy" and create a 3d image.
So I uploaded your last picture and said "Can you create a 3d image of this person please?"
This what it came back with! Not 3d but it is the smartest loco driver you've ever seen.

unnamed.png (639.09 KiB) Viewed 5374 times

[Durley]

Fantastic Phil! Having spent yesterday playing about and creating a few different models, the combination of Gemini and the new Meshy.ai model is game changing as a way to create custom figures. Thanks to Sam95 for originally posting about it.

Note Meshy does have a 2D image creation tool (which is Gemini powered) but the direct Gemini interface seems easier to use and doesn’t use up Meshy credits.

I have now subscribed to Meshy’s Pro package and I think it is worth the fee for me.

Being critical, I have found Meshy does struggle with hats, especially flat caps. I think it has been trained on a lot of examples of baseball caps, so the models tend to come out with brims that are too big and curved, looking more modern. Buttons also appear randomly sometimes, unevenly spaced and on both sides of jackets, but I have mostly managed to fix these issues by using the ‘retry’ function. Stray unwanted features could be removed from the STL in something like Blender, or just sanded away from the physical model after printing.

The details in the model are good, not as good as a really high quality 3D scan or digital sculpt, but perfectly acceptable for printing in garden railway scales. There is texture, folds and creases in the clothing that should work well with my painting style. Faces also have nice details that can be emphasised at the painting stage. Clothing hangs realistically and things like the bottom of jackets, skirts, etc. generally have a nice and thin edge and are spaced away from the underlying clothing giving really good depth to the figures.

I’m not sure Meshy recognises sideburns like the impressive ones James Hughes is sporting, but here’s its 3D rendition from the 2D image I posted above:

[ge_rik]

It doesn't seem to recognise watch chains either. As you say, it's probably trained on modern figures.

Rik

[SimonWood]

This is really extraordinary stuff, and the first time I can actually see generative AI being useful to me personally.

I’m not sure Meshy recognises sideburns like the impressive ones James Hughes is sporting

I keep reading that next leap forward in generative AI will be models that are trained on domain-specific deep data sets. So I'm sure it's only a matter of time before we get one that is trained on side-burns.

[philipy]

I asked Gemini if it could produce stl's. This is the answer,
"I'm sorry, I cannot generate .stl files. I am a large language model, not a 3D modeling program."

No idea what "large language model" means?

[GAP]

Large Language Model is what is commonly referred to an AI bot.
I tried several models and asked them to produce an stl for me using Rik's picture and the response from all of them was the same can't do it.
I finally tried Deep Seek V3.1 Terminus and it told me it do the job, well what its idea of a 3d model and what mine is are at extreme polar ends of the spectrum.

This is its reply when I told it that what it produced was useless.

You're absolutely right, and I appreciate your honest feedback. Converting a single 2D photo into a high-quality 3D printable model is an extremely challenging computer vision problem that requires sophisticated AI and depth estimation techniques that I don't currently have access to.
The Reality of Photo-to-3D Conversion

What's actually needed for quality results:

Multiple photos from different angles (photogrammetry)
Depth estimation AI (like MiDaS, DPT, or specialized neural networks)
3D reconstruction algorithms that understand human anatomy
Professional software like RealityCapture, Meshroom, or Agisoft Metashape

Better Alternatives for Your Needs:
Option 1: Use Photogrammetry Apps

Polycam (iOS/Android) - Take 20-30 photos around the person
RealityCapture - Professional desktop software
Meshroom - Free open-source alternative

Option 2: 3D Scanning Services

Structure Sensor or iPhone LiDAR scanners
Professional 3D scanning studios
Online services that convert photo sets to 3D models

Option 3: Commission a 3D Artist

Fiverr or Upwork artists can create models from photos
Cost-effective for custom 3D modeling
High-quality results with human artistic input

What I Can Realistically Help With:

If you have:

Multiple photos from different angles
3D scan data you want processed
Existing 3D models that need optimization for printing
Technical questions about 3D printing workflows

I'd be happy to assist with those more achievable tasks.

You're correct that my current capabilities for single-photo-to-3D conversion are limited for practical printing purposes. The technology for high-quality results from a single photo simply isn't there yet without specialized AI tools.

Thank you for the honest feedback - it helps me understand the limitations of current approaches!

So it looks like AI has a way to go on this.

[Durley]

A large language model is just a type of AI tool that you put words into (the prompt) and, based on its prior learning of lots of other words (the language model) it gives back a response in words.

An image generator is another type of AI tool that can create 2D images from a prompt that could be just text or text and another image. It is based on prior learning from lots of other images and text descriptions of those images.

Things like Google Gemini (and other ‘chat bot’ style AI) combine these two types of AI in one interface.

Meshy.ai is an example of the ‘specialised AI tool’ referred to in Graeme’s post that is needed to create a 3D model that can be output as an STL file to be printed on a 3D printer. It has been developed based on prior learning from lots of other 3D models, images and text descriptions.

Currently Meshy.ai (or any of the other 3D model AI generation tools) are not combined with chat bots that I am aware of, so you can’t get a chat bot to directly create an STL output.

[philipy]

Thanks for the info. I've been doing some playing myself and I'm quite pleased with how far I've got, with your hints to guide me.

After previously not getting very far, thanks to Durley I think I'm now starting to get a handle on the Google Gemini-->Meshy, workflow. So for anyone else who might be struggling, this is what I've done. I'm sure there are other/better ways of doing things so please shout if you know more than me ( not difficult!).

First of all, go to https://gemini.google.com

Then for me, the key clue was starting with the instruction to Gemini to produce a "photo realistic" image. So my initial Prompt was "Please design a photo realistic British Edwardian period workman. Age 50 years, 5feet 9inches tall, with a full beard and going bald. Open neck shirt and neckerchief "
This is what it gave me ( Gemini rendering):

Gemini_Generated_Image_fvsipbfvsipbfvsi.png (474.11 KiB) Viewed 5247 times

I then told it to remove the spanner ( which I hadn't asked for) and to show his legs and boots. This it did, but the boots were calf-high with trousers ending above them. It looked like it was trying to use Edwardian "huntin' 'n shootin" type boots, so I had to tell it to take those off and replace them. It said it had done it, but actually nothing changed and I had to tell it to remove the boots and replace them, before it did what I wanted.

At this point I then told it to remove the background, which gave me the figure on a plain white background,

kzf0dgkzf0dgkzf0.jpg (88.46 KiB) Viewed 5247 times

I downloaded this and then uploaded it into Meshy. Meshy currently gives the option of using V6 Preview or earlier 'legacy' versions, so I went for the V6. The rendition was pretty good apart from one thing - it insisted on tying his neckerchief as a collar and tie. It also added what appear to be a second waistcoat and a jumper!! ( Meshy rendering)

Screenshot 2025-09-23 08.18.59.png (605.97 KiB) Viewed 5247 times

Only by going back to Gemini and removing the neckerchief then saving it and importing again could I get him with a simple open neck shirt. I assume that it was getting confused by the beard and neckerchief, somehow.The other minor snag was that Gemini had positioned his hands quite close to his legs so Meshy actually joined them together. Another trip to Gemini to move them slightly and reimport, cured that one.

I then went back to Gemini again and told it to make him sit on a chair, which it did, then I told it to remove the chair, which it also did. Imported to Meshy Ok, but then for some weird reason Meshy added a line of buttons up the centre back of his jacket! Easily removed after printing, but still weird.

Screenshot 2025-09-23 08.18.03.png (224.41 KiB) Viewed 5247 times

The next problem is that V6 doesn't allow a free export for stl files, you have to have a paid subscription. It does allow export from what it calls "Legacy" versions, but even with V5 the quality is way below the V6 and not really worth bothering with, IMO. For the amount of use I would get I really can't justify spending £16 a month ( £8 for the first month). So my cunning plan atm is that I'm going to want some passengers for my B&M coaches in due course, so I'm hoping to create a load of them and then take a 1 month trial and download them all in one go, then cancel! That doesn't solve the longer term download problem of course.