VN Ren'Py Catching Heat - AI generated game (technical demo) (development thread)

[deepglugs]

Catching Heat is a adult renpy game/visual novel with lots of AI-generated content. The goal is to produce a entertaining game while pushing myself to both learn AI and push the limits of AI content generating capability. The visual assets of the game are entirely AI generated and I hope to add AI generated script, music and voice to the game later on. AI image generation is a bleeding-edge field of research and the visuals of this game will certainly reflect the lack of capability and maturity of the latest in AI technology. Some assets will be unintentionally scary and of laughable quality. The AI artist is still in heavy development so the generated images ('gens') change with every build of the game. Every week I produce a new build of the demo with the latest gens.

I want to present the game here for feedback on the story within the demo (because that may be the game's only saving grace at this point...).

Game story summary:
Catching Heat is a reverse-seduction harem sandbox game where a succubus and the MC hatch a plan to get a group of girls to try and seduce a boy who is very reluctant to enter into a relationship with any of them. The plan goes sideways when the identity of the MC becomes known and the succubus has to enlist the help of one of the girls to defeat the MC.

Ultimately, I think the technology I'm working on presents some interesting capabilities such as the ability to rapidly produce customized builds with different clothing, scenes, hair, and eventually with different kinks: pregnancy, futanari, etc.

Technical Demo 0.0.20 links/release notes:

You must be registered to see the links

A few of the current "gens" (AI generated images).

Spoiler: Beware: AI generated content

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

Development details you never asked for and don't want:

Spoiler: dev details

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

My plan is to ultimately release an official "technical demo" once quality is barely acceptable. After that, if there is any interest I'll release iterations of the full game. Both the demo and full game I plan on making publicly available for free. Development is expensive, but I'm learning a lot so I get a lot out of it. That said, support and feedback keeps me motivated so feedback of all kinds is welcome!

 

[DuniX]

Porn, the bleeding edge of technology.

 

[deepglugs]

Updated OP with 0.0.11 release info

 

[deepglugs]

Updated OP with 0.0.13 release info. Catching Heat 0.0.14 will be out Sept 4th. I suspect 0.0.14 might be one of the last builds before the final technical demo is released and releases of the full game will begin. This all depends on the quality of the generated images which I expect to be very much improved between 0.0.13 and 0.0.14 given the three additional weeks of training time.

In the meantime, I am training up a new model with the 5th generation neural network architecture. This new architecture should allow for faster training using less VRAM than the current production network/model.

Have fun and let me know what you think of the quality so far!

 

[deepglugs]

Trying out some game banners in blender. What do you think?

 

[Lussuria]

You are a madman. Very unique concept and sounds like a good use of time to get familiar with these AI tools that are becoming available. Recently playing AIDungeon, a GPT-3 based dungeon crawler, I was blown away. To think NVIDIA is working on visual generation and allowing people to mess with it is exciting.

However, from an adult game perspective this most likely wouldn't be something I would be too excited to play from the renders alone. However, I do know that AI generated stories (if I'm interpreting your post correctly, that's what you want to do) can be really funny or have otherworldly feels to them.

I'm curious how long do renders using pix2pix like this take you? I'm just thinking about the magic of an app that essentially combines the real time interaction with GPT-3 of AIDungeon with pix2pix renders being fed info on the GPT-3 created scenes.

The banner is interesting. The heart is off center but the orange icon is interesting. I think you should brighten the font of "Catching Heat." Also, maybe a tagline explaining its an "AI generated game" or something so it might pique people's interest.

Anyway, best of luck, I had to comment out of sheer surprise of what you intend to do.

 

[deepglugs]

However, from an adult game perspective this most likely wouldn't be something I would be too excited to play from the renders alone.

Completely understandable. I'm thinking of adding a slogan to the game: "Play for the story, not the renders". That said, I expect the render quality to improve over time and hopefully approach the quality of traditional rendering techniques. I may have to adopt different techniques, but that's the end-goal.

I do know that AI generated stories (if I'm interpreting your post correctly, that's what you want to do) can be really funny or have otherworldly feels to them.

I've played around with GPT-2 for scene generation. I need to finetune with much more data, but results were indeed funny.

I'm curious how long do renders using pix2pix like this take you?

Pose creating takes a minute or two and the pix2pix network I'm using can generate around 10 images per second. The gen5 network I'm training up now should be at least twice as fast. The pose creation process is human labor at this point, but I have attempted methods to use something like BigGAN and StyleGAN2 to generate the poses from the same image tags. Generating poses is apparently difficult for these networks. I even had stability problems using pix2pix to generate poses from images.

maybe a tagline explaining its an "AI generated game" or something

Great idea! Thanks for the feedback.

 

[deepglugs]

An old attempt at GPT-2 scene generation for the lols

 

[deepglugs]

Sometimes the gens are not bad. To illustrate the model's capability, I can use the same pose and output dozens of different images by changing the attribute tags. Here's just a few character portraits with different clothing and hair attribute tags:
Euryale

Elena

Ceres

 

[Lussuria]

Those look pretty good, especially at that size. I think if I zoomed in I could really see the oddities though. Very impressive technology though.

 

[deepglugs]

Updated with 0.0.14 public build

 

[deepglugs]

Updated OP with 0.0.19 release information. 0.0.19 will be the last weekly build before 0.0.20 which, at the moment, I'm considering the official "technical demo". Using the AI model we are using now, the quality is quiet good and we are moving focus to other things like voice and music generation. With this latest update, we have at least one game character with full AI voice acting (voice pattered after a wonderfully talented human voice actress). We hope to add more voices later, but that largely depends on interest.

If you are looking for a VERY unique game experience, check out 0.0.19 or wait for the full 0.0.20 tech demo which will be a bit better in terms of image and voice quality.

 

[deepglugs]

Updated with final technical demo version 0.0.20.

 

[statess]

I have no idea how to enable NSFW stuff plz help.