From the pictures, it looks like it knows sex pretty well, but is it capable of cunnilingus?

It's doing great job, but there some white-blue dots artefacts i get on V3.0

Seems to be a nice checkpoint! Too bad I use fooocus and it gives bleak images with bright artifact spots. Guess I'll just wait for a version with vae baked

Bro, please train a SDXL Inpaint model with VAE fix! Pleaseeee!

Works very well and doesn't force the girls into a specific style like other models do. So it's more usefull to use! Looking forward to V4 =)

no 12GB version anymore?

='(

Why am I getting a girl that has something like a transparent bottle or condom next to her mouth?

I am getting crap results from this.

I am using Fooocus. With following settings:

Prompt: woman give a blowjob in a bar crowded bathroom

Negative Prompt: watermark, text

Fooocus V2 Expansion: woman give a blowjob in a bar crowded bathroom, intricate, elegant, highly detailed, extremely beautiful, sharp focus, lush romantic, aesthetic, dramatic ambient, cinematic, pristine, background, bright, dynamic, illumination, cute, expressive, perfect, fine detail, pretty, enhanced, complex, elaborate, brilliant, epic, clear, artistic, color, light, striking

Styles: ['Fooocus V2', 'Fooocus Enhance', 'Fooocus Sharp'], Performance: Quality

Resolution: (1152, 896), Sharpness: 2

Guidance Scale: 4, ADM Guidance: (1.5, 0.8, 0.3)

Base Model: sdxxxl_v30.safetensors, Refiner Model: None

Refiner Switch: 0.4, Sampler: dpmpp_2m_sde_gpu

Scheduler: karras, Seed: 1568623116499950957

LoRA [sd_xl_offset_example-lora_1.0.safetensors] weight: 0.1,

is there a model for inpaint?

Ok after my initial surpside how well this model works I did alot of testing with different kind of characters, using controlNets etc. I don't know what you did there, but your model is the most flexible of all that gives good results in all scenarios. I hope you keep building your models in that way!

This model is insane! It works so well in any scenario. Only one small thing to fix, it forces vaginas even on male subjects.

When's the next version coming out? (An estimate of course).

Keep it up, you're amazing!! Thanks for this model.

nice if could improve hands, otherwise just really really good model!

Did you stop uploading the unpruned version?

I would like to test if I can squeeze out a minor quality improvement out of it, compared to fp16.

I think there might be a problem with the VAE you have baked into this 3.0 version. I just ran some images through a basic ComfyUI workflow and all of the images came out very pale and unsaturated. As a test, I stopped using the baked VAE and plugged the standard SDXL VAE into the node and the images came out looking great. Has anyone else complained about this?

Brilliant Model, but when I try and use this as a base to train either a lora or dreambooth I am getting a text encoder error and it won't train. Any idea why this is happening are you using the same pruned model to create you SDXXXL Celeb Loras?

Is there a reason it needs so many steps to be decent? I mean its not bad at 40 but at 50+ its so much better its definitely worth it.

Thing is, using videocombine and animatedif, 10 steps per frame is A LOT.
Even though the results are a bit worse I prefer using a turbo and do 6-7 steps on dpmpp_sde _gpu karass, I save literal hours, exponentially to the amount of work im doing

This is by far the best merge for me, nsfw or not. It gives me better result than JuggXL but it's really slow

Very good quality output, but performance issues? This model seems to be very slow, compared to others SDXL. 1.1s/it average on a 4090 for a 768x1024 image.

This tops juggernaut in my opinion but both are at such a high caliber of quality and now I can't go back to SD 1.5 anymore.

Honey, why is there $8000 missing from savings? Did you buy that used motorcycle you were wanting?

No but I needed two new 24GB GPUs...

Hahahah... oh. Oh. you're not laughing. Wait. You spent HOW MUCH OF THAT on computer shit?!

Well... I needed 256GB of RAM, too.

<:gun shot:>

I use Fooocus, How can i change the steps? For example you wirte "60 STeps recommended"

Really good model, I think it's the best NSFW model among the SDXL finetunes! Compared to Realism Engine the skin quality is slightly lower (less natural and more glossy?). One issue that I noticed is that the nipples are not consistent very often, the left and right have different types. Perhaps it's a general issue with XL nudity, but this particular model gets very close to the top 1.5 models in terms of explicit details!

Very impressive, I hope there will be more updates and improvements.

Not sure as to what I am doing wrong but the pictures always look really good in the preview. But then it always freezes on the last sampling step and the images come out very distorted. Have not had that problem with another checkpoint so far.

The model is just great, the best honestly. Everything works great, but sometimes you need to do the generation without naked breasts. Could you please tell me how to do it, maybe there is some special trigger word for promt or something like that?

What initial resolution would you guys recommend for the best portrait mode results?

If I go with DPM++ 3M SDE Karras, what is the optimal amount of steps to use? Someone has mentioned that this sampler might not be the best choice since it requires an enormous amount of steps to achieve a good result.

How can I prevent text logos from being generated? It keeps making them almost every time. I have all sorts of stuff in negative prompt, typography, text, logo, watermark but it still keeps doing ig

Really great! NSFW model, skin is a little too soft/airbrushed though. A few more fine tunes down the line and this would be the number one or two ranked SDXL NSFW/porn models on civitai. RsmPornXL - Embrace The Suck (Beta) is #1 i think; better prompt adherence and more consistent and reliable vagina.

Creations looks like real porn footage.

Hello, it is an excellent model, do you plan to make a Lightning version?

incredible models

Heads up that I have to run with "--no-half-vae", otherwise SDXXXL errors out at the very end with "NansException: A tensor with all NaNs was produced in VAE." I do not experience this problem with other models. That's using sdxxxl_v30 with no vae override.

Not sure why, but it doesn't produce sex scenes even with prompts and settings copied from existing images here. Only creates nude models with no action. Using it with the latest AUTOMATIC1111 (v1.9.3)

I literally use the same steps and everything about what you put in the examples the exact way and I am not getting nothing like them

Hands down the best nipples out of all the realism models with sharp and detailed texture.

great model

Am I crazy or does this model not know what dicks are? Even when I prompt for a man with a penis, I get a man with a pussy?

Incredible!!

Any plans of using the dataset you used on flux?

So confused here: I have tried both of the latest checkpoints. 3.0, added the VAE in the VAE folder, got haze and anomalies. Moved VAE to checkpoint folder thinking it was a misunderstanding. Got haze, but no anomalies. Used recommended sampling. Same thing happened with 2.0 but no anomalies. I was using pony realism, but keep getting double hip bones. I need some help cuz Ive been trying to find stable checkpoints with loras for weeks.

A lot of watermarking, despite negative prompt.

what is this lora trained on man....

I sure miss this checkpoint