I can guide you with the best practices, tips, and tricks to achieve your goal in your Stable Diffusion training. Take a look at my Patreon, subscriptions for just $5. Exclusive content. Patreon. :)
About:
Well, now let's talk about the model itself.
This v2 of Blowjob POV should have been released at least two weeks ago, but after a conversation with @thegipper, he told me that I was mistaken in the order of the LoRA weights, it was the weight of the UNet being reduced, not the Text Encoder.
This led me back to parameter testing, and also in the meantime there were updates in Kohya that I also needed to adapt to.
The v2 isn't perfect, or much better than the first, but the quality of the images and the photorealism have improved considerably. It's possible to achieve good results without much fuss in the prompts, so much so that I've already got some good photos without even using negative prompts. Now, it's possible to achieve excellent photorealism with practically any base model (considering realistic models).
This version was trained using booru tags, and I tried adding the 'all-fours', 'penis-grab' and 'hand-on-another's-head' poses. It didn't work out very well, but with a bit of luck, it's possible to get some photos with them.
The initial dataset had 250 high-resolution photos enhanced with SD itself, but as I adjusted and eliminated photos that were potentially impairing training, it ended up being only 137, so this also contributed to significantly reducing the possibilities of poses.
I already have the photos for a future v3, which were chosen with much more precision so I won't need to remove them from the dataset later. It's probably going to take a while for me to release this next version because I've been trying to launch an 'Anal Doggystyle POV' for over a month now and still haven't had time to adjust the photos and captions.
The demo images aren't edited, but they are cherry-picked and generated using hires-fix (0.25 ~ 0.5 denoising).
Suggested Settings for Inference:
Model: epiCPhotoGasm (but works with 99% of models)
Positive: penis, fellatio, 1girl, 1boy, blunt-bangs, very-very-high-detailed, nude, trending-on-tumblr, 1 9 9 0, webcam <lora:BjPOV_v2_SD15_by_IsnAI:1>
Negative: lowres, worst-quality, low-quality
Sampler: DPM++ 2M Karras
Steps: 28
CFG: 6
Resolution: 512x768
Note¹: You can get great image variations if you tweak resolution a bit.
Note²: In this version, it's not necessary to reduce weight, it's perfectly possible to generate the images using weight 1 in LoRA (all demos use weight 1), but maybe someone with more patience can discover some interesting things that I didn't have time to test (I'm also not very creative with prompts).
Triggers:
Main: 1girl, 1boy, penis, fellatio, first-person-pov
Tags Used on Training: looking-at-viewer, nude, deepthroat, dark-skin, interracial, all-fours, hands-on-another's-head, hand-on-another's-head, teeth, glasses, sunglasses
Description
First of all, thank you so much for all the downloads, follows, and reactions.
News
Second, I'd like to inform you that I've created two new tiers on my Patreon:
Chat for general support (within my knowledge) for LoRA training and A1111 settings.
A tier for those interested in early access to models, which I haven't finished yet, but that could be used and useful to someone.
Third, I'm considering adding to my Patreon the training configuration files and respective datasets of the models I release here.
About
Well, now let's talk about the model itself.
The v2 of Blowjob POV should have been released at least two weeks ago, but after a conversation with @thegipper, he showed me that I was mistaken in the order of the LoRA weights, and that in fact, it was the weight of the UNet being reduced, not the Text Encoder.
This led me back to parameter testing, and also in the meantime there were updates in Kohya that I also needed to adapt to.
The v2 isn't perfect, or much better than the first, but the quality of the images and the photorealism have improved considerably. It's possible to achieve good results without much fuss in the prompts, so much so that I've already got some good photos without even using negative prompts. Now, it's possible to achieve excellent photorealism with practically any base model (considering realistic models).
This version was trained using booru tags, and I tried adding the 'all-fours', 'penis-grab' and 'hand-on-another's-head' poses. It didn't work out very well, but with a bit of luck, it's possible to get some photos with them.
The initial dataset had 250 high-resolution photos enhanced with SD itself, but as I adjusted and eliminated photos that were potentially impairing training, it ended up being only 137, so this also contributed to significantly reducing the possibilities of poses.
I already have the photos for a possible v3, which were chosen with much more precision so I won't need to remove them from the dataset later. It's probably going to take a while for me to release this next version because I've been trying to launch an 'Anal Doggystyle POV' for over a month now and still haven't had time to adjust the photos and captions.
In this version, it's not necessary to reduce weight, it's perfectly possible to generate the images using weight 1 in LoRA (all demos use weight 1), but maybe someone with more patience can discover some interesting things that I didn't have time to test (I'm also not very creative with prompts).
The demo images of the model aren't edited, but they are cherry-picked and generated using hires-fix (0.25 ~ 0.50 denoising) on A1111.
Suggested Settings for Inference:
Model: lazymixRealAmateur_v40 (but works with almost any)
Positive: penis, fellatio, 1girl, 1boy, blunt-bangs, very-very-high-detailed, nude, trending-on-tumblr, 1 9 9 0, webcam lora:BjPOV_v2_SD15_by_IsnAI:1
Negative: lowres, worst-quality, low-quality
Sampler: DPM++ 2M Karras
Steps: 28
CFG: 6
Resolution: 512x768
Note: You can get more image variations if you also tweak the resolution a bit.
Triggers: Main: 1girl, 1boy, penis, fellatio, first-person-pov
Tags Used on Training: looking-at-viewer, nude, deepthroat, dark-skin, interracial, all-fours, hands-on-another's-head, hand-on-another's-head, teeth, glasses, sunglasses
FAQ
Comments (9)
Buddy, I would like to know how to make the weight of a LoRA more robust, what is the key parameter to achieve that?, and if you could tell me what parameters you focus on when making LoRA models, I have a lot to learn from you, seriously, this is the best LoRA specialized in this, thanks.
hello! what you mean by robust?
also, the most important thing is captions and the dataset
for the 1.5, it's really important that the dataset is pretty concise and has a considerable amount of images, something around 150
I noticed that you can't try to vary the pose too much in 1.5 within the same lora, because it was trained at a low resolution, so complicated poses usually turn out bad. That's why I ended up creating another BJ LoRA for the side angle
regarding the captions, I usually use the interrogate to understand how the Text Encoder perceives certain things, and I use those tags/tokens on training
I did some tests and my best results came from this: <lora:povBlowjob_sd15_by_IsnAI:0.5> (blowjob:1.4), open mouth, wide eyes,
I was able to turn down the lora weight, which is great for customizations, but forced the blowjob tag to make it draw in consistantly.
棒极了。好就好在能做出penis in mouth的效果。
Hi, really sorry for the potentially dumb question, but can this be used for image to image? Or is it only text to image? Quite new to this. Thanks for your time if you could answer!
If you haven't figured it out yet since it's a month old by the time i'm writing this...
it could but if the image is a base depending on the percent difference the quality or actual effect may be weird or nothing or anything in between, therefore use ip-adapter and/or controlnet normal instead.
@DarkSideOfTheForce Hey thanks mate, yeah I'd made that discovery on my own in the past month - usually just warps the face massively. I'll be honest I have no idea what ip-adapter/controlnet normal are but i'll do some research! - appreciate you pointing me in the right direction.
The problem is image to image is designed to adjust whats already there, and if you tell it to adjust forcefully with inpaint then it will make a new image which wont match the existing image. If you inpainted around the lips and turned up the CFG scale to around 7 u might get lucky. Or use inpaint sketch, draw the shaft manually (which will look really bad), but the Lora might work out the relation and draw it in properly. Then move the output from sketch to inpaint and do it again and it can give you some variations to choose from. I think I did this before with good results.
My preferred prompt for this is: <lora:povBlowjob_sd15_by_IsnAI:0.5> (blowjob:1.4), open mouth, wide eyes,