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 190 times

Google Gemini cleaned up 2D version.

IMG_0129.png (557.01 KiB) Viewed 190 times

Meshy 3D model.

IMG_0132.jpeg (126.49 KiB) Viewed 190 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 98 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 50 times

End image ready to upload to Meshy:

IMG_5340.png (552.51 KiB) Viewed 50 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 44 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