discussion AI in adult games. Status of progress for character/asset consistency

[Darkrift411]

I've been waiting for years for AI to get to a point where it can consistently generate the same thing over and over. Lets use the term asset for now. This would be a character, a shirt, a car, a living room, a penis, etc. If I create "dave" and tell AI to make dave standing in the living room by the couch, then ask it to show dave in the kitchen, I expect dave to look the same. His hair style should not suddenly change. The logo on his shirt should not change. His eye color should not change. The couch, now in the background of the shot should not suddenly look different, and the room shape should not have changed. This was something early AI could never do, and made me absolutely HATE AI use in these games.

Aside from that, there are still these oddities in AI generated people that bother me. Usually in the eyes, and often in the torso. Too many muscles, or tendons where they should not be. My brain gets caught up on it and it annoys me because it pulls me from the game. Similar with Daz in girls that are laying on their back with their legs up and you get this odd fold of skin between where the legs connect to the hip. I dislike it, and it bothers me enough to mess with the immersion.

I tried (about 2 years ago?) using LORAs to train AI on a specific "asset". It was a character from a game I was helping with. It did what I thought was decent after some training at showing the assets face, but it changed each render. Even trying to narrow down speifics with the prompt didn't work. And many adult games using AI to gen the images (not talking about ones made with DAZ and then AI just enhances them) have this problem. Most of the time the girls look completely different in each render with only slight similarities.

I know of a lot of devs that suck at making DAZ renders (or HS, etc) but are good at making stories. And we all have seen how often devs partner up to make up for each others limitations. So I have been waiting for a point where AI can consistently generate content like DAZ can.

This post has 2 points. 1, I have seen a few games recently (dont remember the names) that I kinda peeked in on either the game itself, or just the preview images where there was a surprising amount of consistency. Not 100%, but... pretty damned good. So first point is to ask what they are using. Are you a dev who uses AI and is getting very high consistency with assets from one image to the next? Does Jane look exactly like Jane in every image?

Second, I just ran into a video on X that made me feel like this might be getting a lot closer:

You must be registered to see links

I don't think this is there yet, and I would not touch a game made with this, but the fact that this is a feature on a popular AI is kinda cool. Basically, you make a "Reference" and save it. You give it some pictures of what Dave looks like. Some face, some body, etc. Now, when you say @dave it knows to use that reference. I imagine other AI tools, probably even some local models/tools have similar functions either in them already or in the works. This tells me that we are not super far away from being able to use AI to generate some "Assets", feeding those assets back into it and telling it to generate consistent images needed to make these games, or webcomic style stories.

 

[Darkrift411]

Holy shit.... This is exactly what I have been asking for AI to be able to do in DAZ for 2+ years:

You must be registered to see links

I mean... its not perfect, but shit. Being able to load an existing scene in DAZ, and tell AI to move Jane to the kitchen chair and put MC next to her, and change their expressions to X and their clothing to Y... Hell, just the posing and expressions alone might be enough to make most super shitty devs into passing quality game makers. Well, that and make sure none of their models are hovering 2 inches over the ground... but we cant expect everything all at once.

 

[misterfluffums]

I've been practicing with LoRA training for almost a year now, and it still seems the most promising avenue to me.

There are also models that take reference images and / or do faceswapping, and they certainly work to an extent, but a properly trained LoRA should give you higher quality especially in combination with those.

Do you still have your dataset for your LoRA? We could take a look if there are things you could optimize for training.

 

[Darkrift411]

I've been practicing with LoRA training for almost a year now, and it still seems the most promising avenue to me.

There are also models that take reference images and / or do faceswapping, and they certainly work to an extent, but a properly trained LoRA should give you higher quality especially in combination with those.

Do you still have your dataset for your LoRA? We could take a look if there are things you could optimize for training.

No, it was too long ago and the game that was I helping with got abandoned by the dev.

Basically, we took a lot of the best renders from the game, then he gave me the character model and I made some specially for that. Ones like her standing in T pose, or her arms down, from all different angles and I think diff lighting? There was a guide on which types of images worked. And it did a pretty good job on the face, just not consistent. I don't actually remember a ton about the process.

One thing I remember was that I was told there was a point where you hit "overtraining" and the quality starts to drop. I remember seeing that, so I wrote some sort of script that would backup the training data each cycle into a subfolder and run a render against it as an example so we could easily see when it hit that level and stop it and delete anything after it.

 

[misterfluffums]

If the face wasn't consistent, I'm afraid your LoRA was severely undertrained, nowhere close to overtrained

Quality suffering is normal, because there's more information in an image than you could ever fit in a caption. So there will always be random noise which is similar enough in your training pictures that the LoRA gets convinced it's part of the intended behaviour if it faithfully reproduces it.

Bit like your character always standing in front of a white wall, the LoRA will think the white wall is intended to be always present when asked to paint the character (but instead of a white wall it's a pattern in random noise that humans can't see).

So it's a tradeoff between quality loss and faithful rendering. I normally generate the faithful images, then fix the flaws afterwards by only regenerating the parts with flaws. I think you got too scared of quality loss that the model never learned the likeness correctly.

You can also reduce the quality loss by including a normalization dataset that teaches the LoRA "hey you should learn to draw X but don't you dare forget about Y".

But honestly, if you have a daz3d model ready, I would stick to daz3d rendering and maybe layer AI pn that initial render. It's still hard to mimick likeness, so if you have a model already it's more trouble than it's worth.