My attempt at AI spanking photorealism

Welcome to SpankingForum!
Feel free to login or register using the links below.
Please send a message in the shoutbox below for password reset assistance.
Register

shagrath

Guarantor of Guilt
Jan 26, 2025
58
137
1,485
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.
 
These are great!
What model are you using?
I noticed a new spanking lora for Qwen on Civitai, but I think yours is better.
Could you share the lora by any chance?

This website has a ton of stuff, but it's behind a paywall (same as onlyspanking):

In the meantime, here are some Realspankings Network and Punished Brats photos I downloaded to start working on a spanking dataset for a lora model:


For captioning, I use Joycaption:

Then I do some AI assisted batch editing with a python software I vibecoded, it uses LM Studio as a backend to run the LLM.
I normally use it for character Loras, to remove unwanted parts of the captions, but it could be used for a style Lora aswell.
I can share it if you want.
 
Last edited:
You sound way more tech savvy than me. I just use the perchance.org AI generator so have to go with their image base. Got to say it can come up trumps on some imagery but isn't great for a person punishing another person eg. OTK, touching toes etc.
 
Last edited:
You sound way more tech savvy than me. I just use the perchance.org AI generator so have to go with their image base. Got to say it can come up trumps on some imagery but not necessarily isn't great for a person punishing another person eg. OTK, touching toes etc.
You should check out civit.ai
It's meant to be a website for sharing AI models, but you can also use the models through their website to generate images.
Just signup and go to your profile settings to make sure nsfw content is visible and then use the search to find spanking models.
I'm not sure how it works to generate images through the website, because I do it locally on my own PC.
 
You should check out civit.ai
It's meant to be a website for sharing AI models, but you can also use the models through their website to generate images.
Just signup and go to your profile settings to make sure nsfw content is visible and then use the search to find spanking models.
I'm not sure how it works to generate images through the website, because I do it locally on my own PC.
Thanks for that. Can I use spanking images I've got on my pc to help with the AI?

My first hurdle is to give it a go with my VPN on as UK access is blocked.

You must be registered to view images
 
Thanks for that. Can I use spanking images I've got on my pc to help with the AI?
You might be able to use image to image (sometimes called i2i or image2image), to use your image as a starting point.
 
  • Like
Reactions: westpier
These are great!
What model are you using?
I noticed a new spanking lora for Qwen on Civitai, but I think yours is better.
Could you share the lora by any chance?

This website has a ton of stuff, but it's behind a paywall (same as onlyspanking):

In the meantime, here are some Realspankings Network and Punished Brats photos I downloaded to start working on a spanking dataset for a lora model:


For captioning, I use Joycaption:

Then I do some AI assisted batch editing with a python software I vibecoded, it uses LM Studio as a backend to run the LLM.
I normally use it for character Loras, to remove unwanted parts of the captions, but it could be used for a style Lora aswell.
I can share it if you want.

The first pass is created with Qwen image (because I think qwen is much more flexible with obscure concepts) and then it has a second pass through wan 2.2 which refines detail and realism, and it's this combo that works well.

I could share the lora but to be honest it's less impressive than the examples give on. Because i didn't accurately caption the person's state of undress for each training image (underwear on, down to thighs, down to knees, down to ankles, etc.) it will often show the character wearing panties and also a second pair around the knees or something. I'll definitely publicly post the LoRA when it actually has more prompt adherence but I don't want to waste people's time by posting it early.

I would say more of the magic comes from the Qwen/Wan combination workflow, and I can share the json template if you haven't tried that pairing before.

I believe I actually tried Joycaption but it would get some details wrong (it often confuses the left hand for the right hand or it would describe briefs as a thong) then i'd manually fix and it wasn't as automated as I wanted.

Thank you so much for sharing those images! I'm definitely gonna build up my training dataset with those. I can't pull too heavily from realspankings because they tend to cycle through the same 3 spankers and I don't want it to learn a specific person, but I'll definitely incorporate some more. What I would really love is to find more spanking images from vanilla porn productions just because the photography and lighting is cleaner and more professional. But this would obviously be way more time consuming to find and curate, so maybe it's not worth it. The base model should know what high end photography looks like anyway.
 
Got to say it can come up trumps on some imagery but isn't great for a person punishing another person eg. OTK, touching toes etc.

When two characters overlap (especially in poses like OTK) the model often struggles to keep them separate. Diffusion models don't really understand 'Person A' and 'Person B' as distinct objects. They see the entire image as one continuous probability field rather than separate people with boundaries. So when bodies or clothing overlap, features can bleed from one person into the other.

That's why careful captioning during training matters. Without clear textual separation, the model can't reliably assign which features belong to which person. I haven't found the perfect caption structure to fully disentangle the subjects yet, but that's the goal i'm working on.
 
  • Like
Reactions: westpier
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.
That's a great work ! 🌸 What kind of prompt you use to make OTK position ?
 
That's realy nice work, thanks for sharing. It is hard to write the right prompts, like the OTK position, or others as well... so the AI understands.
 
Such beautiful images are just asking for i2v Wan 2.2.
This one I cheated a little, generated in low resolution and then 2x upscale just so it generates faster, and trying out the latest 3 step workflows.
Not a perfect result but it's OK. Might try some native full resolution to see if I can get it to work.

2025-11-07-T230947_00001.mp4

You must be registered to view images
 
You are a king sir. 😭❤️
Just a quick update on this – web scrape is all set. It was able to download ~240,000 spanking images (around 70 GB in total), so not a bad haul overall.

Working on uploading them now, but may take a bit given the size of the dataset! Should hopefully serve as a decent training set.

Here's a small preview:

You must be registered to view images
 
Here's a link to the archive!

None of the file sharing sites I tried could handle this many files, so I created a single 7zip archive and uploaded that instead. Should be relatively easy to unzip, but let me know if you run into any issues doing so.

Cheers
 

Thread starter

shagrath

Guarantor of Guilt
Joined

Thread Information

Title
My attempt at AI spanking photorealism
Prefix
N/A
Forum
Spanking Pictures
Start date
Last reply date
Replies
47