My attempt at AI spanking photorealism

[shagrath]

Here's a link to the archive!

You must be registered to view links and images

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

Excellent stuff! For anyone else who has downloaded the file, don't double click and open the file (it will stall given how many images there are) but instead right click and choose the option to extract to a folder (if you aren't using winrar but a different program this may look different. but here's what my options look like, you want to choose the very bottom option)

You must be registered to view images

So now that i've got a massive collection of images to use, courtesy of Xenon, i'm going to go through and select a diverse, balanced, high quality set of images from the collection and make them training-ready, which means.... manually removing each watermark. 😭

I initially thought i could batch-remove the watermarks using qwen-image-edit but it always modifies the entire image and degrades detail and color accuracy no matter how much i try to tweak the node tree, so unfortunately I believe manually editing is the only way to do it if we really want a superb training dataset.

The way i'm doing this is opening the image in photoshop, selecting only the region of the watermark and doing "content aware fill". Gonna take a bit of time for sure, so if anyone is decent with photoshop and wants to help in this project send me a DM. Otherwise, if you haven't seen me post in this thread in a while don't worry I didn't abandon it, i'm probably just tediously cleaning up images. 🤣

 

[klb1113]

I haven't downloaded the images, but I'm curious: is there any love shown for F/M or is it all mostly F/F or M/F. As an F/M lover, I can tell you we're often left in the dust, including for any AI-generated stuff.

 

[Xenon]

manually removing each watermark.

Noooooo!

You must be registered to view images

Was thinking about this a bit and wanted to propose two possible alternatives to save you time (not to mention your sanity):

1. Does it matter what resolution the images are? I could write a quick script that trims a certain percentage off the top and bottom of the images (i.e., where the watermarks are). Images sizes would end up a bit odd, but if that doesn't matter than this may be a quick and easy solution
   
   
2. There are some AI models that are trained on watermarks specifically (e.g.,
   You must be registered to view links and images
   ) – was able to track down a recently
   You must be registered to view links and images
   that helps make the model pretty easy to use.
   
   I tried it running a few of the images through it to see how it would perform, and the results are actually pretty decent:
   
   
   Spoiler: Click to view watermark removal test 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
   
   There are still some artifacts (for example, Belinda's hand kinda just..disappears in the 4th picture, and her boot is a bit blurry in the 5th), but overall I think it actually did a pretty good job. Might be worth giving it a shot if your GPU can handle it (they claim >1,000 images/minute, though that was using a computer with dual 4090's...I was getting somewhere closer to 100 images/minutes using a single 3080).

Hope this helps!

 

[xaodj]

Ideally, you should use inpainting and just mask the area where the watermark is detected.
That way the rest of the image would not be altered.

 

[jackleson42069]

I haven't looked super closely, maybe it changes some unrelated pixels, but Flux Kontext to my eyes removed watermarks pretty cleanly.
There are a lot of standard nodes to ID a watermark and define the bounding box, you could just crop the picture until the top of the box
There's absolutely no way that this is supposed to be done manually in 2025
The workflow should be more like :
Use Qwen-VL to look at images and help you filter for certain characteristics
Pass all of those to an AI diffusion model to either strip out the watermark or ID bounding box then crop

 

[Jzsman]

Anyone have an idea what this person could be using to generate these?

You must be registered to view links and images

Their videos especially their most recent might be the best ive seen by far

 

[Xenon]

Use Qwen-VL to look at images and help you filter for certain characteristics
Pass all of those to an AI diffusion model to either strip out the watermark or ID bounding box then crop

Exactly. That's more or less what the script I linked in my previous post is setup to do (just uses YOLO instead of Qwen for detection and LaMA for inpainting). Hopefully @shagrath sees these replies before he dedicates too much time to manually editing images.....

 

[jackleson42069]

Exactly. That's more or less what the script I linked in my previous post is setup to do (just uses YOLO instead of Qwen for detection and LaMA for inpainting). Hopefully @shagrath sees these replies before he dedicates too much time to manually editing images.....

Agreed, YOLO is the right path for actually detecting a watermark. I was referring more to the process of bulk processing 230k images - you might want to filter for certain poses, etc. that might be more easily dealt with by Qwen-VL than YOLO. To train a LoRA you really need ~ 50-100 high quality images max, so you could go through a bunch of images in random order, give them some score based on the characteristics you want (e.g. you want some pose, some specific implement, etc.), and pick the top ones.

 

[xaodj]

Anyone have an idea what this person could be using to generate these?

You must be registered to view links and images

Their videos especially their most recent might be the best ive seen by far

These are absolutely amazing!
I'm pretty sure he is using Wan S2V.
Basically you use a starting image and a sound file as reference and it animates the image in sync with the audio.
But it requires a pretty beefy GPU to do this.

 

[shagrath]

The manually removing the watermarks wasn't that bad. I'm quick enough with photoshop that I got it down to 5-15 seconds per image, and if it was a small enough watermark i'd just crop the image instead. Though it was a huge collection of images, it's primarily from 4 studios so I couldn't grab too many images without risking model overfitting due to too many of the same people or interiors.

The captioning is what takes the longest. I thought I could feed it to a vision model for auto-captioning but it was making way too many errors that I think could negatively affect training. It couldn't consistently identify whether a skirt was lifted, or a person was male or female, or what implement they were holding. It had like a 50-70% accuracy for captioning which I wasn't satisfied with.

I vibecoded a very simple webUI where I can select the tags that correspond to the image and it will automatically create and save a txt file for it. All these unique keywords sound counterintuitive, and I thought it would be better to write the caption more like natural prose, but AI is insisting this is the better method for training, so I guess we'll find out. (AI can sometimes be confidently wrong so this will really be an experiment)

You must be registered to view images

 

[jackleson42069]

These are absolutely amazing!
I'm pretty sure he is using Wan S2V.
Basically you use a starting image and a sound file as reference and it animates the image in sync with the audio.
But it requires a pretty beefy GPU to do this.

In one of their videos, they mention that they spanked themselves and dubbed the audio. I think it's regular Wan 2.2 with an audio track.

 

[Sweetpotato]

These were beautiful and sexy I loved them.

 

[Jzsman]

Another channel posting decent AI content

You must be registered to view links and images

Really impressed with how these are turning out

 

[shagrath]

Second try at a non-OTK spanking lora (the OTK pose is complex enough that I think you have to separate that into its own lora)

Prompt: a stunning blonde woman with office slacks lowered to her thighs and in a pink thong bent over a wooden coffee table with her hands braced on the table, a redhead woman in black places her hand across the woman's ass, hand spanking, warm luxurious living room, natural photorealistic professional photography with deep shadows and soft highlights with natural daylight.

Lora:

You must be registered to view images

No Lora:

You must be registered to view images

Prompt: spanking, rear view of a woman in navy blue outfit bent over with her ass sticking out. her plaid skirt is lifted revealing pink lace thong and red ass cheeks. a man partially out of frame places a round wooden paddle directly on her ass cheeks. In front of her is big panel glass windows in a corporate office overlooking a city skyline with natural daylight pouring through the window. cinematically filmed on kodak 35mm with soft highlights.


Lora:

You must be registered to view images

No lora:

You must be registered to view images

The prompt adherence with the lora is definitely more accurate but unfortunately it loses that beautiful realistic style of the base model and I still want to retain that. I'll try to do some more tinkering but just wanted to post an early test to show that i'm still around.

 

[ajw]

Second try at a non-OTK spanking lora (the OTK pose is complex enough that I think you have to separate that into its own lora)

Prompt: a stunning blonde woman with office slacks lowered to her thighs and in a pink thong bent over a wooden coffee table with her hands braced on the table, a redhead woman in black places her hand across the woman's ass, hand spanking, warm luxurious living room, natural photorealistic professional photography with deep shadows and soft highlights with natural daylight.

Lora:

You must be registered to view images

No Lora:

You must be registered to view images

Prompt: spanking, rear view of a woman in navy blue outfit bent over with her ass sticking out. her plaid skirt is lifted revealing pink lace thong and red ass cheeks. a man partially out of frame places a round wooden paddle directly on her ass cheeks. In front of her is big panel glass windows in a corporate office overlooking a city skyline with natural daylight pouring through the window. cinematically filmed on kodak 35mm with soft highlights.


Lora:

You must be registered to view images

No lora:

You must be registered to view images

The prompt adherence with the lora is definitely more accurate but unfortunately it loses that beautiful realistic style of the base model and I still want to retain that. I'll try to do some more tinkering but just wanted to post an early test to show that i'm still around.

I'd like to meet the 2nd blonde. Got her phone number handy?

 

[shagrath]

I'd like to meet the 2nd blonde. Got her phone number handy?

Unfortunately the women trained in the base model are much more attractive than the women typically found in spanking content, and the lora learned that. 😭

 

[ajw]

Unfortunately the women trained in the base model are much more attractive than the women typically found in spanking content, and the lora learned that. 😭

OK. Plan B. Put me in a pic with her. Here's what I look like.

You must be registered to view images

 

[Trollheart]

You're a surfboard?
Hey, I'm not one to judge!

 

[shagrath]

Getting some decent photorealism with Z-image, just need to train longer to get the hand placement and other things right.

You must be registered to view images

You must be registered to view images

 

[jackleson42069]

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?