discussion Viability of AI assets and interest in a High Quality AVN using them

[ZDaGoat]

My main rig is a 9800x3d with a 5090 and 64GB of memory, all tuned and overclocked/undervolted. I also have a second computer with a 5900x and a 3090 or 5070ti I could throw in that to do less intensive stuff if I get to that point. I'm a software developer, a pretty hardcore gamer, and I've been building PCs for over 25 years at this point so I've always had pretty high-end hardware.

I grabbed a few screenshots of some of the initial generations I did for the main girls, they are pretty early and this is without the trained LORAs, but here's a sneak peek:

You must be registered to see attachments

You must be registered to see attachments

First, it's great to find at least a few likeminded people who love to entertain the idea to use AI to create games. I think this will have a lot of future, but right now we are still in the pioneering phase. I hope you guys didn't mean my little game when you were talking about the latest AI slop games out there.

The issue is about consistency. I don't know if a Lora necessarily is the best way. I mean it has good performance once you trained it, but it never trains just the character, it always also trains a certain environment, style, clothing, .... unless you put a lot of work into the training.

With tools like Qwen Image Edit 2511 and maybe the upcoming generation of tools that will even top that, you can really create very good consistency just by working from very well done reference images and prompting the rest you want to have.

Also keep in mind the style of the game you want to make: In your reference images, at least the right one was highly realistic.
Dunno if that is what you really want to do. If so, keep going. As long as you don't dive into Sholi territory as I do, that's fine.

But once you decided for a style, you should try to stick with it. My game is lots of fun and has lots of potential, but it is still struggling from style inconsistencies due to changing models over the time i made it, and on top these video models kind of do what they want, in case you want more than just still images. In the future, maybe evolution cycles for such tools will slow down a bit, and you can resist the temptation to always go with the newest models.

DAZ and 3D are great in terms of consistency, but bad at animations and detailed lifely backgrounds. Maybe the best approach is combining both worlds, which Tieneretsu did for his game Girl Scout Island.

 

[NamoCQ]

A big thing for me is how inconsistent the current images I've seen are in ai games. They don't have to be this way. I would generate a good lora for every character first off. Then I'd generate your scenes separately and inpaint the characters or photoshop the 2 images together after generating. Also scenes with multiple characters are often messed up with loras. inpainting also solves this. Just these little steps and a non generic look would put you miles ahead of the competition.

 

[VNThrowaway]

First, it's great to find at least a few likeminded people who love to entertain the idea to use AI to create games. I think this will have a lot of future, but right now we are still in the pioneering phase. I hope you guys didn't mean my little game when you were talking about the latest AI slop games out there.

The issue is about consistency. I don't know if a Lora necessarily is the best way. I mean it has good performance once you trained it, but it never trains just the character, it always also trains a certain environment, style, clothing, .... unless you put a lot of work into the training.

With tools like Qwen Image Edit 2511 and maybe the upcoming generation of tools that will even top that, you can really create very good consistency just by working from very well done reference images and prompting the rest you want to have.

Also keep in mind the style of the game you want to make: In your reference images, at least the right one was highly realistic.
Dunno if that is what you really want to do. If so, keep going. As long as you don't dive into Sholi territory as I do, that's fine.

But once you decided for a style, you should try to stick with it. My game is lots of fun and has lots of potential, but it is still struggling from style inconsistencies due to changing models over the time i made it, and on top these video models kind of do what they want, in case you want more than just still images. In the future, maybe evolution cycles for such tools will slow down a bit, and you can resist the temptation to always go with the newest models.

DAZ and 3D are great in terms of consistency, but bad at animations and detailed lifely backgrounds. Maybe the best approach is combining both worlds, which Tieneretsu did for his game Girl Scout Island.

I completely understand where you're coming from, and I appreciate the opinion. My approach when I started was definitely that I was going to provide a bunch of info about my world/setting and the characters to an LLM for help creating prompts that I could use to generate image sets for each character that would have consistent styles, locations, their default outfits, and a variety of poses/expressions to train a Lora for each character as a baseline that would help me with a good amount of consistency I could then use to generate/edit assets.

Those images are my early attempts at generating characters, and I was going for a more photo-realism style that I've pretty much decided against at this point. Now I'm wanting a more stylized semi-realism because I am a little hesitant to make characters that look too much like real people. I've also learned a lot about how prompting works and how much control you have over the scene using JSON script with the models, so I don't think I need to go crazy with fine-tuning Loras as long as I have one with facial features/expressions for each character to help with generation/editing.

I haven't posted more, because at this point I've been learning so much that I've been focused on improving my knowledge about the various tools/models. I started out with image generation/editing, but I've also moved into voice/music/sound generation, animation, and I've even been exploring some pretty radical ideas of what might be possible with AI interaction.

I'm also starting to feel like rather than start out by re-imagining an existing (abandoned) game and sinking a bunch of time into a project that ultimately would never be fully "mine" that it's a better idea for me to just decide the kind of game I want to make and move forward with my own creation. I've had lots of ideas I wanted to pursue over the years, and if this is something I am going to sink a significant amount of my free time into then I'd like it to be something I can call my own.

p.s. I can't say anything about your game since I haven't played it, but I have no intention of venturing into the "plus" category. I will say though, that most of the games I've seen with AI animation have serious issues with consistency, to the point that the character's faces don't stay the same and many of them have issues with blurry motion and pixelation or weird proportions. It's definitely not easy to do well, that's why I'm spending so much time learning right now before I make a decision, and why I asked for opinions.

 

[DrPark]

If the quality is up there, I'm sure plenty of people will play it - AI or not. Also promising that you decided to use Astral Lust as a test-run. I really liked that one! Both the creepy setting and because it felt like an actual game.

In regards to ZDaGoat said about consistency: Yes. This is the main issue.
It's not difficult to make single images with AI that are difficult to distinguish from hand-crafted ones.
With some good workflows with ControlNet, LoRAs, etc. it's even reasonable to do so with decent enough control of the pose, character, setting, and style to generate entire image series that are convincing.
Where it really breaks down is consistency within a single scene. Have several pictures of 2 girls in the same scene but interacting (so some stuff changes) and the nightmare begins. The clothing is slightly different, the food on the dinner table is suddenly different dishes, bleed-through of prompting between multiple characters is always a PITA, size consistency where sometimes both are the same height, but then the next image different. Lighting and materials are a bit off. etc. etc.
This is a fundamental issue with AI image generation, that AFAIK nobody was able to truly solve yet - just mitigate (with significant effort).

Standalone each image is good, but as a consistent series? The differences really stand out.

That being said: Don't let that discourage you! If the game is good and the image are "good enough", most people won't care.

 

[dirtyjack74]

I will never intentionally play a game generated by AI prompts. There's no soul in it. I don't care about the "it's a tool" argument, i think it's a terrible thing for something that is meant to be relaxing in our free time, and it being degraded by an AI is just terrible

 

[darknesu]

My main rig is a 9800x3d with a 5090 and 64GB of memory, all tuned and overclocked/undervolted. I also have a second computer with a 5900x and a 3090 or 5070ti I could throw in that to do less intensive stuff if I get to that point. I'm a software developer, a pretty hardcore gamer, and I've been building PCs for over 25 years at this point so I've always had pretty high-end hardware.

I grabbed a few screenshots of some of the initial generations I did for the main girls, they are pretty early and this is without the trained LORAs, but here's a sneak peek:

You must be registered to see attachments

You must be registered to see attachments

Damn 2 years in LC and this is your first thread?

If you don't want it looking like those ai slop and actually sells just don't use raw generated assets right away

These actually look real enough like what