ShareIsLove

New Member
Nov 28, 2022
7
36
So if started experimenting with Automatic1111 and the Dreambooth extension. Do you have any suggestion on how to prepare the images from ORS/GGGB? Should I just add everything or select only images with characters and resize them to 512x512 or keep the original 1280x720 / 1920x1080 and cut them into roughly two/eight images. Also would it help to remove the background so not to muddy the training?
For 1.5 model you have to use 512x512 images.
Use rpaExtract tool to unpack images so you don't have to separate characters from backgrounds by yourself.
Its worth to try to prepare images manually and make few versions of each with different background, for exmaple Ash on black background and exactly same Ash on white background. If image doesn't fully fit in 512x512, you can try to prepare 3 versions of it (or lets say 6 with different backgrounds): fullbody resized down to fit in 512, cut top part without resizing, cut bottom part without resizing. Prepare decent text descriptions for each image so they gonna work with similar prompts later (descriptions can be generated automatically by DB, but it may not work good enough for your case).
Also it might be worth to use some anime focused model such as "anything v3" as a base instead of original SD1.5, and use Danbooru tags in your text descriptions (in that case learn about their tags before you write text descriptions).
Begin with small dataset and just one character, something like 30 images should be enough to get first results and decide if its time to add mode images / characters or if you still need to tweak training parameters.

If you gonna have more questions on this matter send me a private message.
But I won't be able to help you with particular training parameter values as its something you have experiment with (easily can take several days). I'm not an expert in style training and mostly train models based on real humans for deep fake purposes, not artworks.

And to keep it somewhat on topic.. heres my quick fanimation with Stan + Lena.
ORS_SL.gif
hi res mp4:
 

SerHawkes

Engaged Member
Oct 29, 2017
3,323
15,098
If this game will ever be completed, I wish Eva would do DLCs like GGGB afterwards. She can implement cut contents like Ed route, Sidechick Alison (Lena as GF) route, etc.
Hmm, yes clearly because having another creepy old guy would do wonders for this game because everyone so wants to have Ed cheat on his aging wife and then die of a heartattack getting a handy, leaving Lena in charge of the shop. (DOUBTFUL)
 
Dec 17, 2019
170
200
For 1.5 model you have to use 512x512 images.
Use rpaExtract tool to unpack images so you don't have to separate characters from backgrounds by yourself.
Its worth to try to prepare images manually and make few versions of each with different background, for exmaple Ash on black background and exactly same Ash on white background. If image doesn't fully fit in 512x512, you can try to prepare 3 versions of it (or lets say 6 with different backgrounds): fullbody resized down to fit in 512, cut top part without resizing, cut bottom part without resizing. Prepare decent text descriptions for each image so they gonna work with similar prompts later (descriptions can be generated automatically by DB, but it may not work good enough for your case).
Also it might be worth to use some anime focused model such as "anything v3" as a base instead of original SD1.5, and use Danbooru tags in your text descriptions (in that case learn about their tags before you write text descriptions).
Begin with small dataset and just one character, something like 30 images should be enough to get first results and decide if its time to add mode images / characters or if you still need to tweak training parameters.

If you gonna have more questions on this matter send me a private message.
But I won't be able to help you with particular training parameter values as its something you have experiment with (easily can take several days). I'm not an expert in style training and mostly train models based on real humans for deep fake purposes, not artworks.

And to keep it somewhat on topic.. heres my quick fanimation with Stan + Lena.
View attachment 2488768
hi res mp4:
This is really good! Have you considered doing a bunch of these and seeing if Eva would put them in the game?
 
  • Like
Reactions: ffive

LoneVoyage

Member
Nov 22, 2016
154
446
Hmm, yes clearly because having another creepy old guy would do wonders for this game because everyone so wants to have Ed cheat on his aging wife and then die of a heartattack getting a handy, leaving Lena in charge of the shop. (DOUBTFUL)
It was planning to be actually. But then Eva gave up, there were already to many character and we have already Seymour.
 

ShareIsLove

New Member
Nov 28, 2022
7
36
This is really good! Have you considered doing a bunch of these and seeing if Eva would put them in the game?
Its not an easy question.

I'm quite short on time and seems like renpy not so good with realtime animations.
Like, yeah we can playback it naive way simply switching JPGs in a loop, but this exact animation is 48 frames (can be reduced to 12 with some lose in overall quality).
1 such frame in 1920x1080 and 90% JPG quality is ~660kb, 1 animation gonna be around 32mb.
And now imagine that we gonna need to make various animations for every possible character state.
Every combination of tattoos and piercings and clothes and who knows what else we can customize later in the game (hair color? skin color? scars? and so on and so forth). What if more than 1 customizable characters participate in animation? It can go ugly very fast.

Also not sure if it will playback smoothly, especially on mobiles and weak PCs (laptops) as it is extremely unoptimized approach.

I can continue for long... but lets see.
 
4.60 star(s) 336 Votes