AI LORAs for Wildeer's Lara Croft (Development Thread)

Mr-Fox

Well-Known Member
Jan 24, 2020
1,400
3,786
FML I just now realize these are 1152x1152 and I was training them in 512x512. Genius.
---
Medium shots: 001-024
Did you use these as source images for the training of a lora? If so it's not a problem if you have photobuckets enabled.
Then kohya will crop them for you. This can be an issue though if the subject is not centered in the image. For the most control and reassurance that the images are the way you want them, it's best to resize and crop them manually. If your pc can handle it, 768 is better. So if you are using portrait ratio images it would be 512x768. Another problem could be if the resolution ratio is not the same, if resized by kohya since then the image would either be stretched or compressed.
 
  • Red Heart
Reactions: Sepheyer

Sepheyer

Well-Known Member
Dec 21, 2020
1,517
3,568
The third attempt at bodypainted clothes showing promise. The amount of good takes is ~1 in 10? But sometimes goes up to 1 in 2. Naturally, depends what a "good take" is. Bulk of renders are at ~1.0, 1.0 LORA settings, heun as the sampler.

At first I felt the LORA burned the render rather too much and wondered if this is due to anime girls being part of the data set and if I should kick them out for the time being.

However! The LORA was trained on Zovya photoreal v2, and the images that are burned are on Zovya photoreal v1. The moment I started rendering with the native Zovya photoreal v2 then the feeling of burn disappeared.

Zovya photoreal v1:


3RDEdNEWILLUSTRO_illustroV3:

Gotta "sleep on this". But for the time being I feel I need to make changes to the captions to be dead exact about what piece the model is wearing.
 
  • Like
Reactions: AllStep and Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,400
3,786
The third attempt at bodypainted clothes showing promise. The amount of good takes is ~1 in 10? But sometimes goes up to 1 in 2. Naturally, depends what a "good take" is. Bulk of renders are at ~1.0, 1.0 LORA settings, heun as the sampler.

At first I felt the LORA burned the render rather too much and wondered if this is due to anime girls being part of the data set and if I should kick them out for the time being.

However! The LORA was trained on Zovya photoreal v2, and the images that are burned are on Zovya photoreal v1. The moment I started rendering with the native Zovya photoreal v2 then the feeling of burn disappeared.

Zovya photoreal v1:


3RDEdNEWILLUSTRO_illustroV3:

Gotta "sleep on this". But for the time being I feel I need to make changes to the captions to be dead exact about what piece the model is wearing.
Great work Seph. I'm looking forward to the continuation. :)(y)
 
  • Red Heart
Reactions: Sepheyer

Sepheyer

Well-Known Member
Dec 21, 2020
1,517
3,568
I'll post v1 to civitai once the laziness wave passes over me.

Meanwhile, reflecting on where to go with v2 - more strokes or more "smooth", and leaning towards smooth:

 
  • Love
  • Like
Reactions: AllStep and Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,400
3,786
I'm sure you are aware and possibly it was the original inspiration but I thought for "research" to check out those soccer themed bodypaint babe images. The source is sex-worldcup2006 dot com, but you can find a large collection on imagefap.


Example:
1704041309811.png

There are other good examples to "study".
This time it's Brazzers with American football theme. I don't know if this is a one off or a series with more image sets.

1704041580651.png

A few other various examples with other themes and styles

Cosplayer:
1704041806828.png

Sports Illustrated Caroline Wozniacki (tennis) 2016 & Lindsey Vonn (Alpine Skier) 2017

(the whole set)
1704042255545.png 1704042810775.png

1704042692185.png 1704042854899.png
 
  • Red Heart
Reactions: Sepheyer

Sepheyer

Well-Known Member
Dec 21, 2020
1,517
3,568
I'm sure you are aware and possibly it was the original inspiration but I thought for "research" to check out those soccer themed bodypaint babe images. The source is sex-worldcup2006 dot com, but you can find a large collection on imagefap.


Example:
View attachment 3220990

There are other good examples to "study".
This time it's Brazzers with American football theme. I don't know if this is a one off or a series with more image sets.

View attachment 3221006

A few other various examples with other themes and styles

Cosplayer:
View attachment 3221024

Sports Illustrated Caroline Wozniacki (tennis) 2016 & Lindsey Vonn (Alpine Skier) 2017

(the whole set)
View attachment 3221035 View attachment 3221069

View attachment 3221051 View attachment 3221072
Yes, thank you! I have this particular look of what bodypaint should look like (even tho IRL it doesn't look like it), and alas I never found enough images for a comprehensive collection. I kinda like when the look falls apart when close and you right away see the girl is naked. And a lot of IRL photoshoots somehow look too real for me even upclose.

Kinda this is what I would love to find:
 
  • Love
  • Like
Reactions: AllStep and Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,400
3,786
Yes, thank you! I have this particular look of what bodypaint should look like (even tho IRL it doesn't look like it), and alas I never found enough images for a comprehensive collection. I kinda like when the look falls apart when close and you right away see the girl is naked. And a lot of IRL photoshoots somehow look too real for me even upclose.
I thought that the soccer babes were similar to your style and it could be at least good inspiration.
 
  • Red Heart
Reactions: Sepheyer

Sepheyer

Well-Known Member
Dec 21, 2020
1,517
3,568
I thought that the soccer babes were similar to your style and it could be at least good inspiration.
1000%! The soccer babes indeed a complete match. The one thing that stopped me were the inscriptions, etc. I didn't know if I'd be able to isolate these out by captioning the training dataset, so I saved this for the future once I have a lot more experience with LORAs. At this point I am frankly surprised these freaking things I get back from Kohya even work :)
 
  • Like
Reactions: Mr-Fox

Mr-Fox

Well-Known Member
Jan 24, 2020
1,400
3,786
Images without markings:

1704048264881.png 1704047970472.png 1704048003464.png 1704048038021.png 1704048114148.png 1704048147418.png 1704048214778.png
These are a few examples, there are many more in the galleri I linked to and some of them comes from sets that could be tracked down.
 
  • Red Heart
Reactions: Sepheyer

Mr-Fox

Well-Known Member
Jan 24, 2020
1,400
3,786
There are a few lora's on civit the could potentially be used either for study or generate images for the data set.


1704055518561.png


1704055583302.png


1704055642174.png
 
  • Red Heart
Reactions: Sepheyer

Sepheyer

Well-Known Member
Dec 21, 2020
1,517
3,568

Mr-Fox

Well-Known Member
Jan 24, 2020
1,400
3,786
Indeed, I grabbed those LORA before I started. I couldn't really get them to do what I wanted, but that's on me.
I'm trying them out. The first one was really bad and unpredictable. Working on the second one right now. It looks more promising.

Nah, second one was trash. Unpredictable.
 
Last edited:
  • Like
Reactions: Sepheyer

Mr-Fox

Well-Known Member
Jan 24, 2020
1,400
3,786
The third one I linked to.
00037-2275564225.png 00040-2275564225.png 00041-2275564225.png 00046-4023951966.png

meh..

Well actually.. perhaps..
I have experimented with a technique for a while. I'm using an anime cartoon oriented ckpt for the main ckpt and then as a refiner a realistc ckpt but I unly use the cartoon for 0.3-0.4 .
The lora works best ofc on anime cartoon ckpt so this is worth a try.

00058-606959152.png 00059-606959152.png
 
Last edited:
  • Like
Reactions: Sepheyer

Sepheyer

Well-Known Member
Dec 21, 2020
1,517
3,568
The third one I linked to.
View attachment 3221751 View attachment 3221759 View attachment 3221761 View attachment 3221774

meh..

Well actually.. perhaps..
I have experimented with a technique for a while. I'm using an anime cartoon oriented ckpt for the main ckpt and then as a refiner a realistc ckpt but I unly use the cartoon for 0.3-0.4 .
The lora works best ofc on anime cartoon ckpt so this is worth a try.

View attachment 3221807 View attachment 3221809
Alright, I dropped the bunnies here:

I am still on the lazy kick, so if I need to update the post or somesuch, please let me know -- will appreciate a second pair of eyes if you don't mind :)
 
  • Like
Reactions: Mr-Fox