New version includes lots more high res training material. Almost everything is trained at 832x1216 or 1216x832
Uncut has been spun off into its own model https://civarchive.com/models/418871/xpenis-uncut
Some keywords: flaccid penis, erect penis, small flaccid penis, big flaccid penis, small erect penis, big erect penis.
You have to use a VAE, no VAE is baked in. Works best with SDXL base or low CFG for other models
Description
This is a newer updated version with some more diverse and high res training images. Likely will follow with an even better one now that I have a better labeling strategy
FAQ
Comments (21)
All my generations comes out with a strong red tint when using this lora.
No other loras used and I am using the exact same prompts and checkpoint.
Webui A1111.
That's odd. Those generations I posted are all from a1111 and there is nothing particularly red in the source. It's trained on base SDXL. Can you give me an example of which model and prompt so I can reproduce? Or post an example with ther generation data attached. Thanks
@wolffur666456 I mean, I copied the exact prompt, used same model and data as the example pictures.
Love updates. Did you train this model with just penis pics or full body shots including face?
body and penis mostly, I didnt want the faces to get burned in. You can see it can render faces with another model based on the sample images I posted. Just describe the person and the face and add full body or something similar to the prompt. I have another version with even higher res images cooking, I might add some faces in to see. I wonder if there is a good approach to keep the faces in but have it not learn them
I dont think so. I tried your approach and failed at first due to bad prompting. I think it is best to leave the faces out unless the models are really cute.
Does this have any POV photos in its data set?
Not many but there can be more in the next version
@wolffur666456 That'd be awesome! Aside from one Lora someone made of their own dick, there's not many that can do that well haha
So while I want to be as encouraging as possible. It does seem true that 1.3 is far superior. While the art style is retained and artifacts are heavily reduced the subject goes crazy, anatomy goes wild and penises are everywhere they shouldn't be.
I'm using diffusers and lcm, but that shouldn't matter that much considering 1.3b really really was so much more coherent. If you could go back and alter that slightly to keep art style etc and maintain 13bs coherence that would be amazing.
Thanks for making this though, I'll use 13b for making good xl renders then finish them off with 1.5 and the tgirl lora
Thanks for the feedback. I have another version coming with better labels, it's just so much work at SDXL level. The smallest problems with the images make for a lot of artifacts as people complained about the hair patterns etc. They were from some blurry inagea. It's also hard because you have to remove a lot of faces etc or everyone looks like certain faces.
Overall I have made thousands of really great renders with 1.4. With 1.5 the quality is higher and I also made great render but I also use it for inpainting. As in , make the guys using a great full model then use this lora to inpatient the penis. Not that I didn't get good renders with it single shot but like you said, it will put penises anywhere.
I know why it did it. I tried to blend in some blowjob and sex shots in 1.5 for more sexual shots but it super fucked up in that it started to think tongues and other body parts were penises.
I took them all out for 1.6 and found a lot more high res source imagery so the next one should be better.
That being said I'm not removing any versions so feel free to keep using the old ones.
@wolffur666456 nice, good to hear
I do get some good renders quite often. I using them as a base for sd15 so it really does what I need but yea it could be improved a bit. Glad to hear you're not discouraged, you've definitely got the best one available out now. Cheers! : )
@wolffur666456 1.5 really is the best one for inpainting i use it all the time now, dont know how i never thought of it before
@wolffur666456 you mentioned "removing faces". Did you crop the image so that the faces where removed or put a block box over the faces to remove them?
13b gives me really good results while 15e just messes up the whole image!? With exactly the same settings.
Works very well and combines good with woman on the same image. Some said 1.3 may be better, but 1.3 turns every woman into man like appereance and puts the penises on places where they shouldn't be =). So 1.5 is a great upgrade! But penis size is just luck, maybe this can be improved for 1.6? Or even different loras for different sizes, so they dont mix together in some uncontrolable combination.
It seems that since 23/12/2023 it has no effect at all on generated images. Generated images are the same (basicly with pubic hair but no genitalia) with or without Xpenis.
The model has not changed. So something is either wrong with your setup or with where you are generating the images. Are you sure you are calling the Lora?
@wolffur666456 I'm generating online with CIVITAI, with JUGGERNAUT XL Model.
Everything was fine. I was generating nice and convincing photorealistic, and anatomicaly correct results. Then suddenly, on the 23/12/2023, with the same setup, same prompt, same everything, results changed drasticly. Penises simply disapear. Or replaced by some kind of pseudo-vagina.
Is it something caused by the recent censorship striking CIVITAI ? this censorship is not supposed to prevent use of Loras, as I'm aware of... I don't know what happened... And why so suddenly.
v1.5 has improved greatly BUT it still changes the rest of the mans body. For example I render super fat grandpas with white hair. The lora makes them slightly chubby middle-aged with brown hair. Im hoping it will improve and draw only the genitals with affecting the rest of the model.
I take back my comment. I switched sampler and 1.3 works far far better now. Using k_euler now