My attempt at AI spanking photorealism

[funbun]

I don't know much about the technical aspects of text to image diffusion models, so the fact that I tweaked a local model that's 70% close to what i want already.... that bodes pretty well for what's possible a year or two from now, especially if it's in the hands of someone who does know what they are doing. For me unfortunately I think the model is a bit overfitted and I didn't caption my dataset smartly enough. It tends to fix to the same camera angle even when i try prompting otherwise. Probably not enough diversity among my training images. See I don't collect spanking images nearly as much as I do videos, so I didn't have a huge training dataset to build from.

If anyone knows where I can download a huge trove of spanking images from let me know . I mean I know there's this forum, but it seems mostly for video uploads.

But enough yapping here's the results.

You must be registered to view images

You must be registered to view images

You must be registered to view images

You must be registered to view images

You must be registered to view images

You must be registered to view images

Still lots of improvements to be made. Will keep working at it.

Try 'Deviant Art'. Zillions of AI Art work. Also in our preference... spankung, Funbun

 

[shagrath]

Very nice, couple questions:
Is this the new Z-image base model?
Did I read correctly - you training a LoRA on it? Or are you still training on Z-image turbo?

I'm training a lora on the new z-image base model but then i'm using that lora to generate the image with turbo, because it's not only faster way but gives a much higher quality image.

There's still a lot of things to be figured out. If you're using Turbo you have to crank the strength of the lora to 2 or 3.

 

[shagrath]

Z-image base is very strange and kind of broken in terms of lora training, but nevertheless I got some decent results. I'll be trying other models to compare.

You must be registered to view images

You must be registered to view images

You must be registered to view images

You must be registered to view images

You must be registered to view images

You must be registered to view images

You must be registered to view images

You must be registered to view images

 

[klb1113]

May I respectfully request, @shagrath , that it also be able to generate F/M images? The one spanking LoRA out there for Z-Image Turbo (at least from CivitAI) only does M/F. I realize the biggest issue is trying to find appropriate photos to create the LoRA from, though, and understandably there's a lot more M/F or F/F out there than F/M that could serve as training pictures. But CivitAI is sooooo short of decent spanking LoRAs. (Unless there's some other places I should be checking online. I know Huggingface is out there, but it's hard to find stuff on their site without a direct link.)

 

[shagrath]

May I respectfully request, @shagrath , that it also be able to generate F/M images? The one spanking LoRA out there for Z-Image Turbo (at least from CivitAI) only does M/F. I realize the biggest issue is trying to find appropriate photos to create the LoRA from, though, and understandably there's a lot more M/F or F/F out there than F/M that could serve as training pictures. But CivitAI is sooooo short of decent spanking LoRAs. (Unless there's some other places I should be checking online. I know Huggingface is out there, but it's hard to find stuff on their site without a direct link.)

Sent you a dm.

 

[xaodj]

I'm training a lora on the new z-image base model but then i'm using that lora to generate the image with turbo, because it's not only faster way but gives a much higher quality image.

There's still a lot of things to be figured out. If you're using Turbo you have to crank the strength of the lora to 2 or 3.

Have you tried using prodigy_8bit optimizer?
I keep reading that the other ones are not working properly for Z-Image
I'm able to use my ZIB loras on ZIT while keeping strength at 1, but I haven't tried out my spanking dataset yet (I'm focusing on character training atm).

I found the info on this subreddit:

You must be registered to view links and images

PS: any chance you can share your current Z-Image lora in pvt?

 

[shagrath]

PS: any chance you can share your current Z-Image lora in pvt?

Yes, i'm re-running training with prodigy and i'll send you whichever one trained better.

 

[xaodj]

Yes, i'm re-running training with prodigy and i'll send you whichever one trained better.

Great, thanks!
I'm also training one on your old dataset, I'll share it so we can compare results.