Improving old AVNs with generative AI

[Krepax]

Hi

I'm wondering if somebody is working on a software that would be able to take renders from old AVNs and 'improve them' (Better resolution, better detail, better human figures) using generative AI.

I saw a concept on Civitai called 'Render To Real', which is a workflow for ComfyUI (Stable Diffusion) but it's rather awkward to use atm. The results, however are rather good.

Seing how many 'sketch to photo' services there are, I think it should a matter of time before someone does a stand alone package. Considering also how huge the Renpy/VN scene is, I'm also wondering if something along these line already exists.

Consider all the hundreds of old but still enjoyable AVNs, maybe abandoned or for which the author is not interested in doing a 'makeover'. Such a software would breathe second life in many of these older titles.

 

[HELIO CESAR]

Better resolution, better detail

pretty doable

better human figures

not so much
I am not up to date with the AI scene these days, but i belive to achieve this right now the '''''best''''' way would be to use LoRA to work with a image to image approach but i guess it is already what you are using in ComfyUI.

Anyway for the first two cases the are some pretty populars up scaling models for anime out there, one that came to my mind is this:

You must be registered to see the links

 

[Krepax]

I was talking mainly about AVN made with DAZ or similar programs that render "pseudo-realistic" human figures. This is an example of the results of such filters (Render to Real for ComfyUI) on a AVN scene:



The "plastic doll" look typical of so many pre 2020 AVNs is gone. I believe such filters could also be applied to new AVN to improve backgrounds and small details because many of them still re-using the same decade old assets and textures.

 

[witcher234]

it not good when poses are more complex and tends to breakdown.
Also changes faces a lot

 

[Krepax]

Also changes faces a lot

From my limited experience playing with generative AI, it's always possible to "control the strength" to tune the desired effect. Also a certain randomness in the outcome is to be expected but, if enough images are generated, one can always pick the image that fit the most.

I'm not saying such a improvements comes for free or it has to be real time like DLSS and other AI improved upscaling algorithms, but it is something that could help tremendously in the creation of 'makeovers' of old games.

 

[Insomnimaniac Games]

I was talking mainly about AVN made with DAZ or similar programs that render "pseudo-realistic" human figures. This is an example of the results of such filters (Render to Real for ComfyUI) on a AVN scene:

View attachment 4100984

The "plastic doll" look typical of so many pre 2020 AVNs is gone. I believe such filters could also be applied to new AVN to improve backgrounds and small details because many of them still re-using the same decade old assets and textures.

Texture-wise they do look improved, but look at their faces. They lost all emotion, the whole vibe of the render changed. It's the same issue that plagues Pill King, the AI tends to give them thousand yard stares.

 

[Krepax]

Texture-wise they do look improved, but look at their faces. They lost all emotion, the whole vibe of the render changed. It's the same issue that plagues Pill King, the AI tends to give them thousand yard stares.

I've taken a look at Pill King and it doesn't seem bad, as far as graphics are concerned.

Yes, the faces are a bit void of expression, but so are many old games. We all agree that the best way to bring an old AVN up to date would be a complete rework, but how much would it take? Would people be willing to wait months for the renders. Would an author be willing to invest again hundreds if not thousands of houra for something already done?

As far as I can see, reboots usually happens halfway (much to the dismay of the paying patrons), rarely when a game is finished, much less if it's 10 years old...