AI Art generation advice needed

[CrysusPariah2]

I hope this is the right place to post this, if not please let me know the appropriate place.

Me and a few like-minded associates are hoping to create some fan art for a VN we are passionate about.

Professor Amethyst - Love and Jealousy

I am hoping you all could advise us how to proceed or point us in the right direction to achieve matching results.
I've attached some images from the original VN file for reference.
View attachment lacey-sweater-neutral.webp
View attachment lacey-yellowshirt-neutral.webp
View attachment blackskirt&redsweater-neutral.webp
View attachment anna-work-neutral.webp
View attachment l-silverslut-03.webp
View attachment L_affair01_02.webp

 

[eevkyi]

Normally, you would need to train lora models of the characters, which requires either powerful hardware or a third-party service if you don't want the process to take forever, along with dataset preparation and some fine-tuning. However, since the game also uses ai, you might be able to generate reasonably close images with random models + prompts from

You must be registered to see the links

, and some

You must be registered to see the links

.

 

[osanaiko]

I'm not sure that you'd have enough useable material of the characters to do a lora - it would depend on the game content.

My understanding is that you need 100-200 images, hand labelled with the character tags (and clothing or situation as needed), as well as using the same automated tagger as the original model. You'd also need to find a base model that has output that matches the "style" of the game.

Some other ways:

1. Search around to try and find any generated images from the dev still in PNG format - you might be able to recover the INFO block with the model details, diffusion settings and prompt info.

2. similarly, search all the info the dev has written on f95, patreon, discord, etc - look everywhere. maybe they've leaked some info.

3. you could just ask the dev to share info on what model he is using and some prompt examples. might work better if you are high tier supporters.

4. spend some time trawling the model checkpoints/finetunes/merges on civitai, maybe you'll happen on something with the same style / see a sample image that looks similar and can gank the prompt.

 

[animegirl300]

TRAINING LORA IS NOT EASYBUT YOU CAN USE THE REFERENCE ON KITRA AI DIFFUSIO PLUGIN WITH COMFY UI I DID THE SAME TO TRASFORM TO VIDEO IS DIFFIFCULT NO EASY BUT IM GETTING THERE IN NEW TO THIS TOO BUT I SUCESSFULLY INSTALLED WITH THE HELP OF GEMINI NAD COPILOT MICROSFT I KNOW NOTHING OF PROGRAMS AND LIBRAREIS OR DEPENDENCIES BUT IM LEARNING I HAVE INSTALL KITRAI WITH THE AI DIFFUSSION PLUGIN INTEGRATED WITH COMFY UI ..NOW I CAN GENERATE IMAGES WITHOU WIFI AND IF YOU HAVE KWNOLEGE OF PROGRAMING IT WILL BE EASIER FOR YO A FEW PICK FOR DEMOSTRATION

 

[animegirl300]

SO THE PROBLEM ARE THE HANDS ... SO MY PLAN IS MADE A WORKFLOW ON COMGY UI THEN USE IMAGE MADE FOR DAZ STUDIO GENERATE PHYSICS AND MOVEMENT TO SAME CHARACTER MADE SCENE SAME AS KITRA MADE IMAGES SERIES ON DAZ STUDIO THEN ANIMATE SCENE ON KITRA THE POSIBILITIES ARE INSANE I CAN WORKWITH COMFY UI THEN USE A MODEL WORKFLOW AND CUSYTOMIZE TO MY PREFERENCE LIKE I SAI IM OT A PROGRAMER SO IT DIFFICULT FOR ME BUT IF YOU ARE A PROGRAMMER AND HAVE KNWOLEGE OF THIS IT WILL BE EASY FOR YOU IM USING MODEL LANGUAGE TO TEACH ME AND AI MODELS TO LEARN AND INSTALL IF YOU HAVE KWNOLEDGE OF LINUX EVEN BETTER IM USIN WIN 11 AND IS MORE OF A CHANLLENGE TO MADE WORK .. I ASK A AI TO HEP IT DIRECT ME TO WSL INSTALLATIONS SO I HAVE TO REMARK IS FOR WIN 11 YOU ASK ABOUT PRIVACY THEY TELL YOU WHERE TO GO TO NO SHARE DATA WITH SERVERS OF MICROSOFT OR CHROME IS NOT EASY BUT IM GETTING THERE THIS IS THE BETTER EXAMPLE ... YOU CAN USE THE AI TO UNDERSTAND THE PROBLEMS AND CORRECT ISSUES IS NOT GONNA TAKE YOUR JOB AWAY RIGHT AWAY IT WILL HELP YOU TO DO IT EASY TRAIN YOUR OWN AI WHY LEARNID TO USE IT KEEIT FOR YOURSELF WHEN THE TIMES COMES ALL AI CANNOT MEMORIZE DIALOGUES OR CONVERSATION BTU THEY ARE GETTING THERE
Got it — you want me to clean up and rewrite your summary so it’s clear, well-structured, and professional while keeping the meaning intact. Here’s a corrected version of your text:

Corrected Summary
The main challenge I face is handling the hands in my workflow. My plan is to build a pipeline in ComfyUI, then use images created in DAZ Studio to generate physics and movement for the same character. I want to create scenes similar to those made in Krita, using DAZ Studio to produce a series of images, and then animate those scenes in Krita.
The possibilities are incredible: I can work with ComfyUI, design a model workflow, and customize it to my preferences. However, since I am not a programmer, this process is difficult for me. For someone with programming knowledge, especially in Linux, it would be much easier.
I am currently working on Windows 11, which makes the setup more challenging. I’ve asked AI tools to guide me through WSL (Windows Subsystem for Linux) installations. Regarding privacy, I’ve also looked into ways to prevent data sharing with Microsoft or Chrome servers. It’s not easy, but I’m making progress step by step.

 

[mdzone]

Local genning can get you pretty far already. Recent models are decent, though they tend to require at least 12gb of vram and work best with nvidia GPUs.
Im still using illustrious models (with anime/cartoon focus) since Im using an older AMD GPU.

edit: with a checkpoint tuned for realistic images its quite easy to get near the original char already without much finetuning
there are lots of tutorials for stable diffusion out there
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/
https://f95zone.to/threads/your-noob-to-expert-guide-in-making-great-ai-art.256631/

 

[Laetits]

Anna IRL at the Louvre


Lacey IRL in the snow


Lacey having fun at the old granpa's bar
View attachment 17b872ac-c54d-4f37-9dc7-a2c3204094b7.mp4

 

[animegirl300]

Local genning can get you pretty far already. Recent models are decent, though they tend to require at least 12gb of vram and work best with nvidia GPUs.
Im still using illustrious models (with anime/cartoon focus) since Im using an older AMD GPU.

edit: with a checkpoint tuned for realistic images its quite easy to get near the original char already without much finetuning
there are lots of tutorials for stable diffusion out there
https://f95zone.to/threads/stable-diffusion-prompt-sharing-and-learning-thread.146036/
https://f95zone.to/threads/your-noob-to-expert-guide-in-making-great-ai-art.256631/

comfy ui local is good too and better you should try it the models in cividai are free and updated everynow and then

 

[MrFluffums]

I'm not sure that you'd have enough useable material of the characters to do a lora - it would depend on the game content.

My understanding is that you need 100-200 images, hand labelled with the character tags (and clothing or situation as needed), as well as using the same automated tagger as the original model. You'd also need to find a base model that has output that matches the "style" of the game.

Some other ways:

1. Search around to try and find any generated images from the dev still in PNG format - you might be able to recover the INFO block with the model details, diffusion settings and prompt info.

2. similarly, search all the info the dev has written on f95, patreon, discord, etc - look everywhere. maybe they've leaked some info.

3. you could just ask the dev to share info on what model he is using and some prompt examples. might work better if you are high tier supporters.

4. spend some time trawling the model checkpoints/finetunes/merges on civitai, maybe you'll happen on something with the same style / see a sample image that looks similar and can gank the prompt.

Actually, you don't need quite that many images. For my own LoRA's I start off with a single image, that I use with inpainting to create about 6 pictures (1 nude, 1 clothed, 1 bikini, 1 upscaled face happy,1 face sad,1 face angry), and that's already good enough for a first generation LoRA.

OP has plenty images to work with (maybe they just need to add legs to the characters though), it's just going to take a godawful amount of time to prepare a dataset for the 37 characters (I'd estimate 3-4 hours per character). When the training set is done though, any realistic model should basically work.

If CrysusPariah2 is still interested, I could train a LoRA for Lacey with minimal images and show how I did it, so they can do the other 36 characters

 

[CrysusPariah2]

Actually, you don't need quite that many images. For my own LoRA's I start off with a single image, that I use with inpainting to create about 6 pictures (1 nude, 1 clothed, 1 bikini, 1 upscaled face happy,1 face sad,1 face angry), and that's already good enough for a first generation LoRA.

OP has plenty images to work with (maybe they just need to add legs to the characters though), it's just going to take a godawful amount of time to prepare a dataset for the 37 characters (I'd estimate 3-4 hours per character). When the training set is done though, any realistic model should basically work.

If CrysusPariah2 is still interested, I could train a LoRA for Lacey with minimal images and show how I did it, so they can do the other 36 characters

I would be very interested, if you are willing

 

[MrFluffums]

I would be very interested, if you are willing

...okay, this is going to sound silly, but I kinda really underestimated how hard it is to find a pretty realistic style NSFW capable model, because I basically never do realistic I don't suppose you've already found an adequate model that can emulate the style you want right?

I'll have to do some experimentation, I'll let you know when I've found an appropriate base model.

 

[CrysusPariah2]

...okay, this is going to sound silly, but I kinda really underestimated how hard it is to find a pretty realistic style NSFW capable model, because I basically never do realistic I don't suppose you've already found an adequate model that can emulate the style you want right?

I'll have to do some experimentation, I'll let you know when I've found an appropriate base model.

I am told it is ponyRealism_V22MainVAE

 

[MrFluffums]

Okay, I'll have to admit @osanaiko was right and I was wrong. Pure hubris of me of thinking it was this easy with realistic pictures as it is with anime gens

I trained a LoRA for Lacy, but honestly it's not very good. It basically makes her hair explode and mangles her hands and feet even worse than Pony already does. You clearly need to use more pictures than the 8 that I've used, but I honestly don't have the time to prepare all 62 available ones for training.

Anyway, it's good enough to explain what needs to happen. Basically you'd have to upscale as much images you can find from the "images\lacy" folder of LaJ. You then have to create some close ups of her face from those images, some partial body shots and some full body shots. The original backgrounds are transparant, but you need to add lots of colored backgrounds to the images.

When you've created the images, you need to create a text file per image in which you note a trigger word (lacy_laj) followed by everything you don't want to train the model for.

Take the nude image, I've tagged it as: lacy_laj, nude, side view, blue background, hand on hip, full body shot, blurry hair, hand in hair, smiling

Because you've already tagged the image with blue background, the model'll think "I know how to create a blue background" and not accidentally associate it with lacy_laj. Anything that you don't tag (brown hair, green eyes, etc.) it'll start to associate with lacy_laj more and more (which is good, we want Lacy to have consistent hair /eye color).

But honestly, it's going to be a ridiculous amount of work to create a good training data for her, let alone for the other characters.

If you want to test the LoRA you can download it from

You must be registered to see the links

. I've included the pictures and text files as attachment to this post to give you an example of what needs to happen for a LoRA.

I'm sorry for overpromising, I really didn't expect to have such a difference in realistic / unrealistic LoRA's.