Here's another LoRA for a concept that most models do not understand natively. This one works with a wide range of resolutions (I prefer 1152x 896), and is much easier to use than some of my other LoRAs. Baked well with highly curated synthetic data generated over a few days.
EDIT: PONY fixed version released. Really sorry about the blurred faces in the OBSOLETE versions. I neglected this one for way too long. Use ONLY at strength 1.0 (Pony only).
Illustrious is still way better, but Pony version won't produce blurred faces and broken anatomy anymore. I recommend using Adetailer and hires fix.
__________________________________________________________________________________
ILLUSTRIOUS USAGE INFO (v3) -
The illustrious version works infinitely better than the Pony version.
Works great at any resolution and adapts the pose dynamically based on it.
Follow the trigger words, and the prompts in the showcase images to get a good idea on how to use the LoRA.
No obscure trigger words needed this time, just these:
on 1boy, thigh sex, pussy, lying on viewer, on back, head on chest, POV, grinding, thighs together, penis,Use at strength 0.7 to 1.0. I prefer 0.8 for 100% consistency.
__________________________________________________________________________________
If you want to support the creation of more models like this, or you just like what you see (I could really use it, very broke at the moment): ko-fi (｡♥‿♥｡).
__________________________________________________________________________________
PONY ONLY USAGE INFORMATION (NEW VERSION) -
DO NOT USE THE OBSOLETE VERSIONS.
EDIT: RETRAINED - Blurry faces are fixed now.
Use at strength 1.0. Going lower will most likely break the pose or the anatomy. Maybe try 0.8? I recommend 1.0 only.
EDIT: NEW version released. Make sure that you use the trigger words as given.
Cons:
Mention of "penis" can sometimes break the pose slightly (happens rarely). Most Pony models are overtrained on these tags.
Let me know.
How to use (tested prompting structure):
Start the prompt with the trigger words (Try shuffling them for variety):]\
REFER TO SHOWCASE IMAGES FOR PROMPTING STYLE FOR PONY.
Trigger words:
thigh sex, p0vth1ghj0b, lying on person, upside down, grinding, lying, on chest, male POV, legs together,__________________________________________________________________________________
In case of any doubts or problems, feel free to ask or comment. I can only fix it if you let me know about the problems.
Description
Very consistent, no facial distortions, and flexible with minimal style loss.
Refer to showcase images for better idea of the prompting structure.
FAQ
Comments (14)
great work ! would love to see female fingering male !
the faces are blurry no matter what i do, detailed face and blurry face tags do nothing, sadge
The Pony one, right? Illustrious version will not have this problem. I'll update the pony version in a few days.
I did some weird experiment with the older dataset which backfired. I absolutely hope it's not the illustrious version though.
@nottheninja yeah its the pony version, good to know yore on it cause other than the faces this works really great
lol... I've tried so long a couple of weeks ago... and gave it another go yesterday... I was sure I was doing something wrong, since I couldn't even reproduce the example images... it's not even the about the pose ... I just need to find a solution to problems... it has been my job for years after all lol ... but I guess I will wait for your "all new and shiny" fix ;)
@pillowaifu @AvonN Sorry about the trouble. I used an Adetailer with high denoise to fix the faces in the earlier models. The new Pony version should work great now.
@nottheninja no worries - we all make mistakes and can learn from them :)
I seem to still be having problems haha... maybe it is me this time?
My first remix of an example image of yours but switched to real dream checkpoint just to experiment removed the dick - some follow up generations with some test prompts of my own drew the dick one time of tens of images (different prompts). My no-change remix test of a different pic of your examples (batch size of 2), got me a similar picture on the first image, a completely blurred patch on the face on the second. any idea why?
@AvonN What's the strength you're using the LoRA at? Also, you're using the onsite generator, right? I just tested with Real Dream, using this LoRA at strength 1.0, and a batch size of 2. Looks normal. Let me test further. (EDIT: Added one gen with Real Dream that you can refer to in the gallery)
Make sure that you have the "Mature content" toggle turned on while generating. Some of the prompts don't have the tag "penis" in them. Add it for more consistency maybe?
The showcase images are generated on tensor with a latent hires pass at 0.4 denoise, and one pass of Adetailer. If you need help with doing this, just reach out in DMs. (I prefer the tensor onsite generator way more compared to the one here).
@nottheninja I had started wtih a weight of 7.5, but then saw the description and upped to 1... did another remix-no-change of the previous image: second image blurred (mature content active).
RealDream remix-no-change removes the dick lol...
This is without ADDetailer - I only use the onsite workflow for images I find ok. (it seems a bit wonky to me especially with hands ;) ).
Well - civitai is my home of choice at the moment - so I generate here ;)
I have recently noticed that even with random gens prompt changes are flacky sometimes... I've already had a (unintended :P) PC reboot today... might try a cache deletion later on....
Noticed the remix added the safe_pos/neg_pos keywords on the realDream gen - these mess up remixes from experience.... removed them for a retry and realDream remix seems ok afterwards.
played around with it some more... it doesn't seem to like added loras or embeds very much :P
@AvonN Yeah, it kinda seems to not play too well with some LoRAs. Also, lowering the strength than 0.9 will break the pose most of the time. The dataset for this is very rare, and makes it less flexible as a result. Maybe one day I'll find a better way to fix the Pony version. Meanwhile, the illustrious version works so good that I may not prioritize it.
@nottheninja Hmm... can't you use generated images for the dataset, if it works for you? (me = little knowledge of training models :P)
might give the illustrous version a go, once I explore those models ;)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.