• Sign up or login, and you'll have full access to opportunities of forum.

Hornet1ba - My 3d Renderings

Go to CruxDreams.com
Thanks for all the likes! For what i should have mention is a work in progress 3d-AI creation
with a new workflow that is promising but need refinement as the missing nails.

Also the fact that the custom head morph created by mp5stab is completely replaced by a random stranger is not the best use of AI for now.
 
Hornet1ba said:
Thanks for all the likes! For what i should have mention is a work in progress 3d-AI creation
with a new workflow that is promising but need refinement as the missing nails.

Also the fact that the custom head morph created by mp5stab is completely replaced by a random stranger is not the best use of AI for now.
Click to expand...

I found this an interesting challenge, so I gave it a shot after @Hornet1ba generously gave me permission to do so.

Converting a Daz3D render to a photorealistic image is easy enough. But doing so while preserving the likeness of a face is a different matter, for which I could see no clear solution.

My first attempt was to use as many controlnet images as possible to capture the essential traits from the original. Normally, I'd use just the lineart model as stacking up controlnet images tend to deteriorate the image quality. And because lineart is generally not enough to preserve the likeness of a face, I also tried to use canny and reference models which I don't use very often.

Also, I experimented with normal and colour, hoping they might help transfer finer details and colours of the face respectively. It didn't work well which made me doubt the usefulness of those particular control net models.

As expected, using that many controlnet images together made it look rather unrealistic. So I masked the face area to preserve its content then inpainted the other area using just lineart and depth, which resulted as below:

01535-20230708165007.jpg

Note that I didn't bother to depict the nails and the blood because it's not the main challenge in this case. I already showed that it's doable in another render of mine before.

I wasn't happy with the result entirely (especially the skin tone, and the way the cross looks) so I decided to try it again using a different approach.

This time I started as usual, trying to find the right tone and lighting first using only the essential controlnet images. The result looked more realistic than before but the character didn't resemble the original as much, which was quite expected.

So, I just inpainted the face area using a similar controlnet setup as used in the first experiment. It restored the resemblance somewhat but had the side effect of reintroducing the "3D render like" feel to the face. So I ran the whole image through a series of img2img processes using reduced reference model influence (and very low denoise strength, of course). It mitigated the issue but didn't eliminate it entirely.

It seems that it's a matter of choosing the right spot between the perfect likeness and the Daz3D-flavoured uncanny valley. So, below is the point I chose between those two extremes:

01577-20230709070333.jpg

Again, converting a 3D render to a photorealistic image itself isn't a difficult task, and it's quite easy to render images in the likeness of a given character if there's a sufficient number of reference photos.

But converting a single 3D render into a photorealistic image without destroying the likeness of the character can be a more difficult problem. So I took it as a challenge and a chance to learn control models that I haven't used much. And I'm glad that I did since I learned much from this experiment :)

Again, thanks @Hornet1ba for allowing me to use his original work!
 
Last edited:
fallenmystic said:
I found this an interesting challenge, so I gave it a shot after @Hornet1ba generously permitted me to do so.

Converting a Daz3D render to a photorealistic image is easy enough. But doing so while preserving the likeness of a face is a different matter, for which I could see no clear solution.

My first attempt was to use as many controlnet images as possible to capture the essential traits from the original. Normally, I'd use just the lineart model as stacking up controlnet images tend to deteriorate the image quality. And because lineart is generally not enough to preserve the likeness of a face, I also tried to use canny and reference models which I don't use very often.

Also, I experimented with normal and colour, hoping they might help transfer finer details and colours of the face respectively. It didn't work well which made me doubt the usefulness of those particular control net models.

As expected, using that many controlnet images together made it look rather unrealistic. So I masked the face area to preserve its content then inpainted the other area using just lineart and depth, which resulted as below:

View attachment 1322101

Note that I didn't bother to depict the nails and the blood because it's not the main challenge in this case. I already showed that it's doable in another render of mine before.

I wasn't happy with the result entirely (especially the skin tone, and the way the cross looks) so I decided to try it again using a different approach.

This time I started as usual, trying to find the right tone and lighting first using only the essential controlnet images. The result looked more realistic than before but the character didn't resemble the original as much, which was quite expected.

So, I just inpainted the face area using a similar controlnet setup as used in the first experiment. It restored the resemblance somewhat but had the side effect of reintroducing the "3D render like" feel to the face. So I ran it through a series of img2img processes using reduced reference model influence. It mitigated the issue but didn't eliminate it entirely.

It seems that it's a matter of choosing the right spot between the perfect likeness and the Daz3D-looking uncanny valley. So, below is the point I chose between those two extremes:

View attachment 1322104

Again, converting a 3D render to a photorealistic image itself isn't a difficult task, and it's quite easy to render images in the likeness of a given character if there's a sufficient number of reference photos.

But converting a single 3D render into a photorealistic image without destroying the likeness of the character can be a more difficult problem. So I took it as a challenge and a chance to learn control models that I haven't used much. And I'm glad that I did since I learned much from this experiment :)

Again, thanks @Hornet1ba for allowing me to use his original work!
Click to expand...
All around a learning experience and thanks for the detailed explanatio!
 
The Daz rendering\Stable Diffusion (AI) hybrid experiment continue, now with a body and face morph created by our own mp5stab in the Daz Render and 2 version modified by AI where i used my first model trained with about 10 images of the morph created by mp5stab.

If you don't understand the last part of my description it's OK :cool:
Emily on a A shape cross5.jpg01292-946925730-a nude girl crucified to a wood post,screaming_.5, black curly pony tail h...jpg01302-1640746587-a nude girl crucified to a wood post,emily20-550_1.4, black curly pony ta...jpg
 
Wragg said:
I'm no expert in all this, but I believe the feet are the original render, the blending between the Stable Diffusion and the Daz needs a bit of colour adjustment, maybe. ;)
Click to expand...
You are correct: i did mask the feet and hands of the Daz render when i let Stable Diffusion replace the character as Stable diffusion will most of the time create
even worse feet and hands than the Daz 3d mesh also the feet of the Genesis 8 figures are kind of bland.
This AI tech is promising but not without flaw!
 
MSTGMA said:
The feet are still a bit...uh...:p
Click to expand...
Hornet1ba said:
You are correct: i did mask the feet and hands of the Daz render when i let Stable Diffusion replace the character as Stable diffusion will most of the time create
even worse feet and hands than the Daz 3d mesh also the feet of the Genesis 8 figures are kind of bland.
This AI tech is promising but not without flaw!
Click to expand...

Aw, c'mon, guys. Hornet's doin' great feet (or is that great feat?).

Everything's relative. (And you should see my relatives' feet! Ugh!) For example, Hornet's footsies are spectacular relative to, oh, say, a lemur's.

human.jpg(Hornet's creation)

lemur.jpg(A lemur)
 
poem21045 said:
Aw, c'mon, guys. Hornet's doin' great feet (or is that great feat?).

Everything's relative. (And you should see my relatives' feet! Ugh!) For example, Hornet's footsies are spectacular relative to, oh, say, a lemur's.

View attachment 1323695(Hornet's creation)

View attachment 1323696(A lemur)
Click to expand...
I did interpret that as a constructive criticism and i did look into some way to improve the feet: there is some add on to add realism and vascularity to the feet and stable diffusion is having less problem producing realistic feet when it zoom into the feet.
01622-3565514523-2 feet nailed to a wood post,hd,highly detailed,realistic,canon 5d.jpgG8 feet2.jpg
 
Hornet1ba said:
I did interpret that as a constructive criticism and i did look into some way to improve the feet: there is some add on to add realism and vascularity to the feet and stable diffusion is having less problem producing realistic feet when it zoom into the feet.
View attachment 1323722View attachment 1323723
Click to expand...
Tree would like to note spikes with large heads would hold her to the cross instead of using wood washers and, after she succumbs to the cross, they can be used again on the next condemned person...
 
Good pillory. The woman's hands are securely securely captured without excessive pinching. She is bent forward as her head is fastened in its port. The count is impressed with the devise's construction. To celebrate her successful confinement. He positions himself behind her and opens his trousers. He ignores her gasp as his hard shaft enters her...
stock pillory 046.jpg
 
You must log in or register to reply here.
Back
Top Bottom