FappXL
This model is a experiment of how many concepts I can fit into a single rank 16 LOHA model. This is the SDXL version of https://civarchive.com/models/129350/fapp-multi-concept
Samplers:
The following samplers worked best for me
Restart
DMP++ 3M SDE Karras
DMP++ 2M SDE Karras
The DMP++ 3M SDE and DMP++ 2M SDE samplers (non Karras) don't work.
Concepts (triggers in Brackets):
High quality Pussy (pussy, innie pussy)
Cowgirl position (sitting on man, penis penetrating pussy)
Cowgirl from behind (sitting on man, penis penetrating pussy, from behind)
Ass Poses (from behind, sitting from behind, on all fours)
Feet (feet, soles)
Legs Up poses (legs up,feet up)
Sideways poses (laying on side, on side)
Penis (penis, erect penis, flaccid penis, penis bulge)
Blowjobs (sucking a mans penis, male pov, penis in mouth)
POV doggy ( from behind, male pov, penis penetrating pussy)
POV missionary (on her back, spreading legs, male POV, penis penetrating pussy)
Creampies (cum in pussy, creampie)
Buttplugs (buttplug, heart shaped buttplug)
Tgirls (need to also use my TgirlsXL Lora)
This LORA wa trained on base SDXL, so it should work with all Checkpoints. However the Sex concepts work best with SDXXXL
Take a look at the samples for inspiration
Description
First version for SDXL. Uses the same Dataset as the 1.5 version
FAQ
Comments (37)
you have a very creative name. i bet if i search for it in russian, hungarian or polish telephone books, i will find it ...xDDD
really bad maybe 1 in 100 works
also every woman gets a penis surgery
which checkpoint are you using?
@ffjggrtbjibv JUGGERNAUT XL
I have this problem also, really bad renders. SDXXXL.
@saytomepls Which sampler are you using? For example "DMP++ 2M SDE" gives terrbile results. "Restart", "DMP++ 2M SDE Karras" or "DMP++ 3M SDE Karras" work well
@ffjggrtbjibv DPM++ 3M SDE Karras works reasonably well, restart rather poorly. Previously on the restart I tried, but now it is bearable. Thanks for help.
the image quality went up after I set the steps above 40
Unfortunately it struggles pretty hard with the base XL model, which is unfortunate as the base model still seems to put out the most accurate LORA characters. BJs are bad, but that's a bit expected with base XL. Feet are also quite bad though.
Switched to Albedo to try more feet and unfortunately not any better there. Feet look the same as without the LORA loaded and it's giving me a lot of distortion in bodies.
I like the idea though. Keep at it.
Saw the other thread so I'm trying Restart on Albedo now. 50/50 for really distorted gens still. I feel like the base model does a better job with feet without the distortions.
Not here to complain btw, hopefully the feedback helps.
for some reason, I'm getting bad results with this. when copying the prompts of others for tests I've noticed it's like my lora is missing the mark just barely, or starting to bake in certain elements and guarantees distortion on a generation when using it. I don't know why. are we not using the automatic11 webui anymore or something?
does not seem to have any effect on a turbo SDXL checkpoint
Can't seem to get it to work, no matter the model(all SDXL obviously)
Which sampler are you using? For example "DMP++ 2M SDE" gives terrbile results. "Restart", "DMP++ 2M SDE Karras" or "DMP++ 3M SDE Karras" work well and the steps need to be above 40 if not using Turbo models
@arslan2012 will have to try again but I think I was using turbo models? Not sure. Thanks for the suggestion! I will try again.
i`m going to try. and make sure u use Karras
A question about training, did you put all the datasets together and train them all at once? Or did you use some method where each concept was trained separately and then combined together?
All at once
Not sure why, I even copied existing prompts and modified them a bit, then used the recommended checkpoints. But 90% of the pictures are a mutated monsters party...
EDIT: I got more consistent results with betterThanHentaiXXXL checkpoint, but still a lot of mutations.
Same as the others. The prompts, even with the listed checkpoint does not produce these images.
Believe it or not, the most utility I get from this lora is not so much its intended use, but rather the fact that it improves the geometry of anatomy. I sometimes also use it at low strength (like around .30-.50) for the skin texture, and other general features, it produces. It is very realistic or "lifelike". (Then at some points, it becomes too realistic or "lifelike" enough that it actually becomes uglier than if I toned it down or even didn't use it at all and actually prefer it the other way, so this must be done with some oversight. But still useful.)
I've reached this conclusion after having been using this for a long time now.
So, yes, the collateral effects are positive.
Great idea! While SDXXXL is quite good, it often adds watermarks and the skin it makes is too smooth. Using this lora with RealVisXL (for example) gives fantastic results! Really good skin + all the NSFW capabilities if needed. At 1.0 weight it produces nightmares but 0.3-0.4 works well by adding to the main model and not fighting it. I also used RealismEngineXL from SDXXXL author, it works fine too.
@rkfg For the record, I've been using this thing with models like the Super Mario Merge (at first) and RealitiesEdgeXL LCM Turbo (later) and ultraspice xl turbo (a bit) and now I just introduced Dreamshaper XL Lightning as the main (all previous models I listed are XL Turbo, since it's highly convenient) and I expect it to work here just as well when I need it
Using a lora at a really low strength to bring in some skin texture. Clever! I'll have to play around with that.
@lucidzachary473 I play around with low values. Even 0.25-0.10 produce interesting results. What I mean by "geometry of anatomy" is things like, the shape of the vulva being generated better, or the ass getting better shapes (though not always), or - importantly - the model somehow now understanding certain poses better like kneeling and thus generating less anatomical nonsense as a result.
I use this one but also - as I suspected - nsfw pov all in one for this sort of effect more recently and I often mix both while playing with the weights.
I guess this is probably not exclusive to these loras; it might be that it's a general property of how they are trained (and on what). Loras are made so that the model can do one new thing right, but they affect generation in a myriad other ways, often harmless, but sometimes actually beneficial as here, even though that was never the intended result. Hope that clarifies.
i rarely use this above 0.25 strength. It's a strong, strong LoRA, folks, which is why you see a lot of complaints. Turn it down, way down, and it'll behave. Don't try to use this as the single source of your NSFW scene, it's just not well-suited for that. This is better for assisting with anatomical details.
I downloaded this model and the file name is FappXL.safetensors. but when I go to LoRA menu and load this file, it is loading up a transgirl Lora. <lora:tgirl_stage2:1>
@kenu169 correct, but that's just the erroneous name of the LoRA internally. The author has another LoRA that is an actual Tgirl LoRA and this was originally meant to be related to that in some way. But in practice this LoRA has never produced a Tgirl for me.
I think this is the wrong lora, it is the tgirls, can you please check. Thank you
It's the correct one Civitai chnages the name to the one thta was used while training. I forgot to change the name in the config after training the Tgirl lora
Can you link your workflow?
Works great, just have to keep the strength low (<0.30)
Feels like I just got trolled.
The internal name of this lora is "tgirls" and it only produces women with penis.
tgirls are girls with penises...
@DanBrody holy fuck that was my point special guy.
the lora is not named TGIRLS on civitAI but it is in Forge.
I can't believe you needed that explained to you smh
@DanBrody These are men. Just feminine