Basically "assisted fellatio", but one of the girls mostly cares about encouraging the other girl, to complete the task at hand.
Prodigy-Trained on Pony XL V6, but I have more luck using pony merges, than the checkpoint itself. Includes activation words for PoV and look a little bit more afar/voyeuristic (Voy_). Examples per Checkpoint.
I'm not good at prompting, but had most luck with Euler A, DPM2, DPM adaptive. Prompting different people is kinda hard, but using character names (if the checkpoints understands them) or character Loras might make your life easier. Strength 0.7 is often enough but depends heavily on checkpoint.
Might be (fantasy) sexual counseling, corruption, or just two girls having fun, encouraging each other to dare new things. Could probably also used for other scenarios, but remember the rules and don't post inappropriate stuff here.
Description
FAQ
Comments (33)
This lora look hot. Is there any chance for SD 1.5 version?
I tried before (without much success). I might give it another try, though.
Still not exactly happy with it, but you can give it a try: https://civitai.com/models/301884/sex-tutor-fellatio-sd15
@ragnaroook Thank you ^^
Why clip skip 2?
A habit from SD1.5 - where anime seemed always to need clip skip 2. Let's see if I can persuade my GPU to run this again (for clip skip1), without crashing...
@ragnaroook I don't think clip skip 1/2 is a thing in sdxl. For some reason I can't find actual info about it, but in automatic1111 changing clip skip makes no difference and in kohya changing clip skip with sdxl gives out a warning about sdxl not expecting a clip skip value or something.
Edit: Just looked at sdxl lora metadata and clip_skip is always set to "None", so I'm pretty sure it just isn't a thing, and people are just mistaken when recommending clip skip 2 like pony diffusion author.
@ragnaroook That's because of the NovelAI model which was heavily trained with CLIP skip 2 and it became a popular model to merge with so it made sense for LORAs to be trained with CLIP skip 2 as well.
What doesn't make sense is training loras with CLIP skip 2 when the model was not.
Description of this model:
"Prodigy-Trained on Pony XL V6"
Description of PonyXL:
"Make sure you load this model with clip skip 2 (or -2 in some software), otherwise you will be getting low quality blobs."
What makes no sense is people randomly complaining about clip skip 2 when the model clearly works. Use whatever works for you and maybe learn to read descriptions.
@hhks People writing descriptions for models are also often just making guesses and passing it as facts though. Nobody actually knows much for sure on here. Clip skip is not even a valid parameter for sdxl, the blobs have nothing to do with it, that's just some error that randomly happens on pony diffusion when certain prompts cause it to fail at the first steps.
@wewewew Taking into account that the model creator recommends Clip skip 2, and that ponyXL has failed everytime I generate with Clip skip 1, but works fine with Clip skip 2, it tells me that Clip skip is in fact a 'valid paramater for sdxl'. I am very sorry, but I will choose to believe the actual model creator and follow their advice, instead of someone who 'thinks clip skip is not a thing in sdxl', 'can't find actual info about it' and then is suddenly 'pretty sure it just isn't a thing'.
@hhks Well, I am interested in what environment your situation happens, to verify this. I've been pretty clear how my opinion changed from not being sure to being sure and you seem to just ignore that. What platform (automatic/comfy/other) and version does clip skip make a difference in?
What it could be, is that sdxl in general requires clip skip 2, and any other value fails, and so every platform that properly integrates sdxl automatically sets it to 2 and ignores user input. In effect, it would be fair to say clip skip isn't a valid parameter for sdxl in such a case. But anyway, I just want to figure it out, it's very weird that this and lots of other information around AI usage is impossible to find, and everyone's just throwing out guesses because it's easy to come to conclusions like "so and so makes the output better" when a lot of it is very inconclusive and non-deterministic.
I am not an expert in the field of neural networks, but I declare that the Pony model works perfectly in Fooocus and FooocusRuined, despite the fact that there is no way to explicitly specify the " clip skip". (Automatic translation into English)
This looks good, but does anyone know how to change clip skip in fooocus?
No clue, but In A1111 it doesn't seem to make a difference(for this model), whether I set it to clip skip 1 or 2. However, I'll give it a new run anyways...
Can you make one for Anal?
If I find the time to collect samples I might give it a try.
such bad prompt ... like SD15
thats SDXL !!!
The prompts I come up with, work even worse for SD15, so if you come up with something good, give us a glimpse.
its not that easy ... if you want automatic tags ^^
but you can try
BooruDatasetTagManager - semi automatic
and also add easy manual tags for groups of images
@sevenof9247 How would automatically throwing Booru tags at a prompt be any different than SD15?!
I like this but I hope for a v2, this version tends to give characters huge lips, and odd faces. or at least it did for me...
Thanks for the feedback. I don't know if I can train this better, without getting a whole bunch of new source material (current version relied on cartoon-y examples, which might explain the faces being all over the place and exaggerated lips), but I might give it some fine-tuning with better prompts some day (describing the lips and oddities in greater detail might help the Lora to "get" that these are more optional)
In the meantime you might try to prompt for "thin lips" or (depending on the generator you use) try inpainting, to replace the odd bits.
what means "VOY"?
As in "voyeuristic" - a little more afar than pov
I don't have a lot of luck with this Lora 1 out of every 5 comes out as a candidate and of those only a couple look "okay"
I've mostly used it with external generator A1111, not the on-site generator, but you might try to use it at lower strengths (depending on the Checkpoint sometime 0.5 is enough, but 0.7 seems to be a better bet). Also I had more luck with checkpoints originating from Pony than Pony itself. My best bet was AutismMix with the DPO-lora built in, but that's not available as well. Perhaps try to add in the DPO-Lora at low strength?
@ragnaroook I appreciate the advice. I've really fallen in love with Pony and Pony Realism. Will try with autism. Thank you for the advice.
@seruko862 Good luck and thanks for giving my Lora a try.
I did a test-run with the on-site generators checkpoints, https://civitai.com/posts/3336069 , and yeah it doesn't look that great on average.
@ragnaroook there are deffo some stand out's though. Good work
i love this "voy" idea. i've noticed trying to achieve the same effect by using variations of "side view/from above" for example only achieves that desired effect once every 5 roles. looking for something in between POV and the rough generation of "side view" and voy achieves that well... so well in fact that this LORA's being used to take advantage of the "voy" token in contexts other than the sex tutor. do you have any plans in the future to build more SDXL pony LORAs based on this unique perspective? i haven't seen it anywhere else, and using "voy" and "voyeuristic" doesn't seem to prompt well
I tried to do a "photobomb" lora, where sex happens more in the background (say a group shot or wedding photo of know characters where interesting stuff happens in the background) , but couldn't get it to work in the slightest - I guess on the one hand because SD is focused on putting your prompt into the focus, on the other hand I didn't have much source material.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.