VN Ren'Py Our Red String [Ch. 12 Full] [Eva Kiss]

[ShareIsLove]

Just my 2 cents on AI theme.
Midjourney is fun and games but Stable Diffusion is much better for real tasks such as turning sketches into illustrations with pose/features/color/light preservation.

In general, to get decent results heres must be an artist behind the tool as someone must serve as art director (decide on art style, train AI models for chosen style, prepare base images for AI to work with, put everything together, manually fix issues such as hands).
Also to make most of it its preferable to have decent hardware (ideally nvidia with 24gb, but can go with 8gb or even free Google Colab) and being able to train your own models based on particular style and characters (both can be your own or mix of many) to have high quality and consistent results across the project.
Issues like hands and genitals can be solved to some extent by model training too, but for artist it will be easier to fix such issues manually.
For writers who has stories but lack of drawing skills the best approach is to take photos and screenshots from videos as base material, make lots of AI generations based on them, mix best parts from different generations together and overpaint manually where needed (if you have hands and genitals in your references its not hard at all to stylyze them manually on top of AI generation).

 

[UsagiTripleSix]

In general, to get decent results heres must be an artist behind the tool as someone must serve as art director (decide on art style, train AI models for chosen style, prepare base images for AI to work with, put everything together, manually fix issues such as hands).

Sounds right. I was blown away at how easy it was to just make beautiful topless women...only to discover how hard it was to do something more complicated than that--or to try to get a specific art style. I'm already looking at training to be able to reproduce the same character more than once, and I assume training might be needed to get clothing to be the same from image to image.

There's also a lot of struggling with prompts to get elements like lighting and facial expressions to do what you want it to do.

So, while I do think we'll see it as an alternative to the typical DAZ/Honey Select/Real Porn methods of creating art for games, would-be creators will quickly find that they're going to need to sink in probably the same amount of time and effort to master to the tools, if they want anything better than a couple of individual images.

Also to make most of it its preferable to have decent hardware (ideally nvidia with 24gb, but can go with 8gb or even free Google Colab) and being able to train your own models based on particular style and characters (both can be your own or mix of many) to have high quality and consistent results across the project.

I've been doing mine locally with an 8GB nVidia card. Seems to be the bare minimum VRAM required, and I already find I struggle trying to do images at reasonable resolutions or with my training attempts.

 

[jimul18]

View attachment 2473474 View attachment 2473605
Just my 2 cents on AI theme.
Midjourney is fun and games but Stable Diffusion is much better for real tasks such as turning sketches into illustrations with pose/features/color/light preservation.

In general, to get decent results heres must be an artist behind the tool as someone must serve as art director (decide on art style, train AI models for chosen style, prepare base images for AI to work with, put everything together, manually fix issues such as hands).
Also to make most of it its preferable to have decent hardware (ideally nvidia with 24gb, but can go with 8gb or even free Google Colab) and being able to train your own models based on particular style and characters (both can be your own or mix of many) to have high quality and consistent results across the project.
Issues like hands and genitals can be solved to some extent by model training too, but for artist it will be easier to fix such issues manually.
For writers who has stories but lack of drawing skills the best approach is to take photos and screenshots from videos as base material, make lots of AI generations based on them, mix best parts from different generations together and overpaint manually where needed (if you have hands and genitals in your references its not hard at all to stylyze them manually on top of AI generation).

Is there a ways/tool to improve or train the models based for example on the input images (i.e. the specific porn stills that E.K. used) and output images (i.e. the finished illustrations by E.K.). For GGGB the inspiration was pretty obvious at times. I don't know enough about how to create g.u.i. and train a model to for example be able to input a new porn still and tell the image generator to transform it into the illustration style similar to E.K?

This would be pretty amazing...

View attachment 2474799

 

[ShareIsLove]

Made some more for fun.

So, while I do think we'll see it as an alternative to the typical DAZ/Honey Select/Real Porn methods of creating art for games, would-be creators will quickly find that they're going to need to sink in probably the same amount of time and effort to master to the tools, if they want anything better than a couple of individual images.

I would not consider AI as a full blown alternative to anything really (only to manual overpainting / rotoscoping to some extent).
To get best results out of it you should combine it with other tech such as 3D or real photo/video references. Or your own drawings.
So if you need lots of images with same character and clothes, it still will be reasonable to make at least simple 3D model to use as a reference for generations (img2img + controlnet). By "make" I assume using something like Daz / Character Creator / MakeHuman / MetaHuman + Blender for posing and clothing (also heres many good blender addons for character creation and free models around).
It may look like even more complicated approach than just use 3D, but you gonna have much better results than with AI alone or low effort 3D alone. So more like "next level" than "simple solution".

I've been doing mine locally with an 8GB nVidia card. Seems to be the bare minimum VRAM required, and I already find I struggle trying to do images at reasonable resolutions or with my training attempts.

With all optimizations enabled you should be able to do Dreambooth training with 8gb. In earlier days lower limit was around 11gb but afaik they pushed it to 8 quite some time ago.

You must be registered to see the links

is worth to take a look, I'm not sure if its able to run on low vram cards tho.
I tried to fine-tune SD1.5 with face of real human and results was very good, almost perfect. Training on characters should work just fine.

input a new porn still and tell the image generator to transform it into the illustration style similar to E.K?

If you have capable GPU, to do so you need

You must be registered to see the links

with Dreambooth addon to train original Stable Diffusion model (1.5 version is a good starting point) with EK style (StableTuner from the link above may give you better results, but for babysteps Dreambooth should be okay as it has many tutorials).
Check videos about Dreambooth on youtube to learn about the process.
In short, you gonna need to prepare images with desired style (so grab all of GGGB images and cut them into 512x512 pieces with characters), put them into same folder and run training. Once its done you can either "redraw" any existing photo/image with EK style or generate completely new images with it.

 

[UsagiTripleSix]

Is there a ways/tool to improve or train the models based for example on the input images (i.e. the specific porn stills that E.K. used) and output images (i.e. the finished illustrations by E.K.). For GGGB the inspiration was pretty obvious at times. I don't know enough about how to create g.u.i. and train a model to for example be able to input a new porn still and tell the image generator to transform it into the illustration style similar to E.K?

Well, there’s no magic bullet that will do all of it for you... Though I think Eva is in a good position to take some advantage of AI, since it’s clear that reference images are being traced to get the foundation of the art. An AI model could probably be trained on samples of ORS graphics to make it mimic the art style (likely need one for characters and one for backgrounds).

I‘ve heard LoRAs are good ways to go about doing that, but I have yet to try it myself. I’ve made a few attempts at training Textual Inversions to get the AI to reproduce things like faces, but I’m still learning how to do it effectively. I think training an art style would be a more complicated project. It’s been done, though. You want a Studio Ghibli look? You can already get that.

To get best results out of it you should combine it with other tech such as 3D or real photo/video references. Or your own drawings.
So if you need lots of images with same character and clothes, it still will be reasonable to make at least simple 3D model to use as a reference for generations (img2img + controlnet). By "make" I assume using something like Daz / Character Creator / MakeHuman / MetaHuman + Blender for posing and clothing (also heres many good blender addons for character creation and free models around).
It may look like even more complicated approach than just use 3D, but you gonna have much better results than with AI alone or low effort 3D alone. So more like "next level" than "simple solution".

That’s actually been my starting point. I’ve already gotten fairly good with Honey Select, so I’ve tried my hand at using my work from there as the basis for AI images. Some things work, some don’t. Getting basic poses is fine. Trying to reproduce characters via ControlNet alone (without training), not so much. Plenty of other things I have yet to try, though.

And everything I do goes through Photoshop. I can’t imagine ever getting perfect output from HS or Automatic1111, so there’s always touch-up work and merging of multiple images to get a better result.

 

[kisskissbambam]

Oh, are we doing AI fanart? I've been working on figuring out Automatic1111 and making decent strides. This tech is really going to change things...

View attachment 2473091

WOW!!! You should make your own version of ORS and have this Cindy as lead... Amazing skills

 

[Night Hacker]

WOW!!! You should make your own version of ORS and have this Cindy as lead... Amazing skills

What skills?!

 

[kisskissbambam]

What skills?!

Is there none? IDK, I'm not a dev

 

[UsagiTripleSix]

Is there none? IDK, I'm not a dev

There is some. Got to be skilled enough to read and watch tutorials, get the software properly installed, understand how to use it, read up on effective positive and negative prompt construction, figure out what plug-ins you need, how to install them, what settings to use, etc., etc.

So artists won’t be losing their jobs to AI, necessarily. But they may lose their jobs to people who take the time to learn how to use AI.

And it probably would be fun to try more ORS or GGGB images in Automatic1111. Unlike my game, where everything was HS models, EK’s illustrations have real human proportions that don’t appear to mess up the AI nearly as much.

 

[Night Hacker]

Is there none? IDK, I'm not a dev

Niether are they. That's just more AI crap. No skills required.

 

[balvenie1401]

Niether are they. That's just more AI crap. No skills required.

Not really an issue when the resault is leaps and bounds beyond 99% of the artist/devs on here.. Would love to see writers/devs getting proficient with AI and not having the render count being a constant excuse for failing to deliver, how wonderful that would be

 

[Night Hacker]

Not really an issue when the resault is leaps and bounds beyond 99% of the artist/devs on here.. Would love to see writers/devs getting proficient with AI and not having the render count being a constant excuse for failing to deliver, how wonderful that would be

Nah... I prefer to do my own work, not rely on programs. There's a bit of a copyright grey area with that as well as AI doesn't "create" images like that. It get pre-existing images and is basically like an AI photoshop.

I'll ALWAYS prefer people and their skill than just some smuck and so called "AI".

 

[Samuel Hidayat]

Is there none? IDK, I'm not a dev

A proficient artist can make good AI arts.

A random schmuck can't because they lack basic drawing and artistic knowledge. That's why you see a lot of bland AI arts with gross anatomical mistakes and uncanny valley aesthetic—something good artists will never proudly publish.

 

[seaoflove]

Not really an issue when the resault is leaps and bounds beyond 99% of the artist/devs on here.. Would love to see writers/devs getting proficient with AI and not having the render count being a constant excuse for failing to deliver, how wonderful that would be

AI has severe limitations. There are visual art AI and there are writing AI.

Writing AI will not replace human writers. It might help polish things up after something is already written, or offer some suggestions or help proofread, but the human touch is always very much needed.

As for visual art; It may be helpful for some developers as is writing AI. Perhaps it could generate some usable background scenes for a game like this. That may be a fair use. However, AI visual art would still need a base image to work off of if actually generating characters. If it were the only thing that was used it may wind up drawing the same character 50 different ways and it would not be consistent.

Both types of AI tend to be web based, and also tend to be heavily in favor of censorship.

You want ChatGPT to help write an erotic scene? Good luck!

You want midjourney to draw you porn? Good luck!

Both have censorship programs built in. On top of that artists are filing a huge class action lawsuit against companies that use AI to scrape images from the web without anyone opting in. So, these companies are going to be facing major litigation in the future.

I wouldn't count on AI making a game any time soon. Even if such a powerful AI existed so it could write perfectly, code perfectly, and render perfectly it would still need a ton of prompting and training to get a desired effect.

Human made art is still superior to the machines.

 

[UsagiTripleSix]

Thought of a interesting project for someone AI savvy with way too much time on their hands: Thought it could be interesting to train a model/LoRA on the style of ORS and then try applying the new coloring style to the GGGB assets; give the game a visual upgrade.

It’d probably be an enormous undertaking, though, but from what I’ve heard, Stable Diffusion is pretty good when it comes to doing things like coloring linework.

 

[Defenestrator.for.traitor]

A question what is the best way of construct the book Ian wrote?

 

[seaoflove]

Thought of a interesting project for someone AI savvy with way too much time on their hands: Thought it could be interesting to train a model/LoRA on the style of ORS and then try applying the new coloring style to the GGGB assets; give the game a visual upgrade.

It’d probably be an enormous undertaking, though, but from what I’ve heard, Stable Diffusion is pretty good when it comes to doing things like coloring linework.

That might be something cool to see. I could see AI helping to make a mod like that, but to create an entire game from scratch not so much.

 

[ffive]

Guys, no offense but could you maybe take the AI talk to a more suitable thread? It's quite enough to see these talk points rehashed whenever a game with ai cg tag pops up, which is like 3-5 a week nowadays

 

[Samuel Hidayat]

A question what is the best way of construct the book Ian wrote?

Spoiler: Book points

You don't have permission to view the spoiler content. Log in or register now.

Given it seems to be a problem, an easy and practical guide to achieve high results, and even the best one, in writing Ian's book.

You just have to choose: Vengeance, Relativistic, Anti-Hero, Doom, Metaphysical, Sacrifice.

In this way you'll achieve 8 points (aka perfect score) with a Historical book, and 7 points with a Fantasy or Sci-Fi book. To reach max points with those two you have to choose: for the Fantasy one Romantic or Crude, for the Sci-Fi one Chosen One. That's it.

To make it even simpler, just choose these for max score:
Fantasy: Vengeance, Relativistic, Anti-Hero, Doom, Romantic, Sacrifice.
Sci-Fi: Chosen One, Relativistic, Anti-Hero, Doom, Metaphysical, Sacrifice.
Historical: Vengeance, Relativistic, Anti-Hero, Doom, Metaphysical, Sacrifice.

If instead you want more freedom in choices, these alternatives will all give you the perfect score anyway:
Fantasy: Vengeance/Call Of Duty, Villain/Relativistic, Anti-Hero, Doom, Romantic/Crude, Victory/Sacrifice.
Sci-Fi: Chosen One, Villain/Relativistic, Trickster/Anti-Hero, Treason/Doom, Metaphysical, Sacrifice/Defeat.
Historical: Vengeance/Call Of Duty, Relativistic, Sage/Anti-Hero, Fight/Doom, Romantic/Metaphysical, Sacrifice.

 

[Jason Grimm]

Can someone post what Eva said in the most recent patreon post: Gillian's new look
Please