wait it isn't?? wow, then it's very impressive. I thought sdxl was incapable to do some things. It's actually better than Ponys... very impressive, mate. What a great work.

2 is for anime. just test both

Yeah, it really is!!

Is LCM +Exponential/Karras a single scheduler ? Fooocus has all 3 listed as separate schedulers, and I have a Dev option for 'Forced Overwrite of Sampling Step'.
4 steps instead of 30 sounds like Lightning or Hyper SD though ? For me, they never give as a good a result as the 30 steps 'Speed' setting, or 40 steps 'Quality' setting of Fooocus.
I pronbably shouldn't touch it, should I :-)

Ближние нормальные, на средней дальности и чуть дальше губы сливаются с зубами

@coyotte вообще не встречал чекпоинты где на расстоянии не шакалит зубы и пальцы

face/hand детейлерами, не всегда, но в целом помгает.

у LUSTIFY V20 встречал проблемы на средней дистанции, но больше было похоже на кривой CFG или когда излишний макияж.

короче, проблемы в целом стандартные для всех, решается фиксом промпта, поиграться с cfg и детейлерами обмазаться)

@coyotte кстати, может знаешь в чем проблема. ОЧЕНЬ редко, но иногда из-за чего-то получаются мужские руки у женщин 

Пользуясь случаем, поинтересуюсь, может есть способ "убрать мужика и\или его достоинство" из кадра)) Вроде даже в негатив всё накидал, но настырный, всегда сует куда и когда не просят)) А то сидит девушка, делает селфи, а тут откуда не возьмись "волшебная штука" прирастает снизу))

Ну и почти все существующие модели даже оригинал SDXL не понимают разницы между "SMIRK" и "HAPPY" , обязательно надо скалить зубы хотя это простая улыбка, насчет рук я добавляю в негативный промт Male hand,male fingers если модель нужна только женщина, а так да почти все модели (даже оригинал SDXL) пихают мешают части тел)))

@coyotte Its not. Actually incomatible with character loras, and celebrity likeness is poor. If you can make progress in that field plus more beuty(just check the most beautiful checkpoints what i wrote), than this would be the pony of SDXL, simly the best. Until then its just for porn. Its the best porn checkpoint, but it could be more.

@coyotte Okey dokey

I got curious, what prompts did you use that you couldn't achieve beauty? :/

Lustify v2 lighting can do some really stunning images I will soon upload them to my gallery, and I can create the most realistic nsfw images with my own trained models indistinguishable if it's AI or real. This is by far the best NSFW model I have used so far. Yes, I have tried to merge it with Dream Shaper, Realviz increases the quality, but the down side is it loses its nsfw capability a bit.

What do you mean by beauty? What's your goal with a checkpoint?

Beauty is extremely subjective in the realm of ai checkpoints because each checkpoint has it's own specific goal it's trying to achieve.

If you mean beauty as in generating women that have beautiful faces, then it's a matter of your prompting.

I've generated some gorgeous women even when I didn't intend to when using Lustify. That shouldn't be difficult.

@lisaxoy677229 While i agree character loras doesn't give you good likeness but the checkpoint itself is definitely good. Also try training loras for this checkpoint

@chrisss1 Indeed, training a lora on this checkpoint gave a very good results, especially for a character lora.

@coyotte Ah too bad. Then SDXL it is! Are you training localy or here with Buzz? Would be willing to donate for some Flux experiments. I have too much fun with it to let the potential go to waste! But Anyways, great job and I will patiently wait for any update you do =)

@WhatTheGuy forget about flux. Nobody is doing it because you need literally thousands of dollars and flux has serious issues (finetunes crumbling at x steps). It's just not worth it.

@ElectricDreams Too bad, then let's wait. There are models that are de-distilled and one that is retrained for better lora training. Maybe in the future there is new tech than improved Flux training... Or even a flux.2 Model ;P

@WhatTheGuy Nvidia's Sana model is coming soon.

@ElectricDreams yeah examples look great! But who knows how trainable it will be after all the problems with SD3 and Flux ;P

Looks like v2 is the best one for inpainting

If you don't like DMD2, just use the Lightning lora with the original v4. SDE Turbo sampler/scheduler looks good. Lightning also tends to push the style more towards glamour

@coyotte mmm yes, LCM and Karras. i don't see anybody else complaining about it so I'm probably making something wrong. I need to try exponential scheduler. Nevermind. Don't bother. Anyways, great job mate. Base model is truly an advanced model. I hope you keep working on it. Cheers.

@ElectricDreams I had to increase the steps to 30+ to get good details

use ultimate SD upscale

How does it compare to Pony Realism?

Could you elaborate? How is it better than the main model?

settings

amateur photo, 1woman, medium breasts, naked, sitting in the back seat of a limousine, busy parking lot, fingering herself, expression of pleasure, thicc,
Negative prompt: (worst quality:2), (low quality:2), (normal quality:2), (lowres, error), username, watermark, ugly face, mutated hands, (blurry), blurry face, blurry iris, poorly Rendered face, poorly drawn face, poor facial details, poorly drawn hands, poorly rendered hands, low resolution, Images cut out at the top, left, right, bottom, bad composition, logo, banner, bad teeth, mutated body parts, blurry image, disfigured, oversaturated, (bad anatomy), deformed body features
Steps: 40, Sampler: DPM++ 2M SDE, Schedule type: Karras, CFG scale: 7.5, Seed: 3858338787, Size: 1024x1024, Model hash: 34caf7f4fc, Model: lustifySDXLNSFW_v40

@coyotte Thanks!

Same here... trying different VAEs and stilll wrong

@coyotte Bursting with excitment

"penis" is a VERY hard prompt.

The genital is mostly "more normal siced" if you're not using "penis" in any prompt

Try a soft weight debuff like (penis:.4) in negative?

@coyotte Thank you for the reply, that was useful, looking at the positives rather than the negatives seems to be the key. Negatives seem to do nothing.
I'm still investigating. But I think I figured it out.

I'm using large tables of wildcards, and large batches, so it's tough to pin down. But some combos cause it more than others.
I'm doing suggestive but non-sexual all-female nudity, and the dick pops up in the combination "group of" or "other women around her" and a mention of sexual poses or specific female anatomy.
Mentions of specific anatomy + 2 or more people in the scenes = random dicks.
I'm going to try to move the references further apart in the prompt and see if that improves things.

It's probably just what you said, the shear amount of porn. The model is just thinking "(there should probably be a dick in their!)" and I can't really blame it.

Amazing model btw: In my humble opinion, the very best NSWF out their right now, thank you.

@coyotte Found another STRONG one. The perspective of "view from slightly above" or "from slightly above" creates a "male pov" with lower torso and/or dick. Not all the time but more frequently than normal.

I can't really figure out how to fix that though without removing Lustify's amazingly good POV shots.
I'm also intentionally creating these male pov shots and the model is great at it!

I think the #1 improvement I could suggest is the complete elimination of men/dicks in all girl scenes.
perhaps with just a keyword that calls out girl-girl, or lesbian and includes no man. Currently it makes a lot of chicks with dicks or adds men. I'm sure some people want that but perhaps it can be called out in negatives or positives.

Overall though I really love the model. I do primarily "professionally" shot, non-porn, fashion with explicit nudity and people in the background and the model is the best around at that! You even have my camera, Canon 5D!

Can't wait for V5!
I'm happy to help in anyway I can.

That's what I'm saying!!!

@gghostg1400 Haha, this creator has literally created a monster! I didn’t know that LCM could fix hands so well. I’m so in love—I even left pony SDXL behind

I feel a bit left behind, but what is DMD2 and how can it be used with Forge? Thanks for enlighting me :)

@johnjohnson Is the lustify 4.0 dmd2 model, you can generate using LCM sampling with only 7 step to get good image, i am using forge too, so no problem, as long as you have LCM sampling in your ui, and set the CFG 1 to use this model

@ColorWolve Wow THANKS so much!

@johnjohnson you are welcome

What settings are you guys using for DMD2?
i.e. Sampling method, Scheduler type, Sampling Steps, CFG scale and also, which Upscaler is optimal for this DMD2?

thank you.

@makaveli_313"I use the LCM sampler with the scheduler set to automatic. I set the steps to 7 and the CFG scale to 1. I generate images at 1280x1280 resolution directly, without using any hires fix. but i do use adetailer to fix the face

@ColorWolve thank you.

@ColorWolve I completely agree!! I recently tried out "Tame Pony" which is now in my number 2 spot. This one knocked it out of the top spot and it's gonna sit there for while till something better comes up. I have been on Lustify since version 1 and didn't like what it did with body shape, so I went on to Pony and months later I was like "hmm, wonder what the newer versions are like" and booooooooy!!!

@gghostg1400 well, I can say that Pony does a great job with anime style, but when it comes to real people, the body often looks too plastic to me. Lustify, on the other hand, is a game-changer. The fine-tuning on this model is incredible, and the body looks realistic—just like a real photo.

@ColorWolve a follow-up question please, do you use any styles or if you have any prompt suggestion to get realistic image, amatuer kind of shot not 'airbrushed' look on the skin.
Also, for Adetailer its face_yolov8s.pt ? or anything else for the faces.
Thanks again!

@makaveli_313 I start my prompt with 'Shot on Sony A7' well, you can use others camera model, and without adding terms like 'masterpiece' or 'best quality.' In the adetailer inpainting settings, I set the denoising strength to 2 to keep the face looking realistic. I also adjust the 'Inpaint only masked' setting to 200, since the default of 32 brings it too close and makes the skin appear unnatural. I use the LCM sampler with the schedule set to automatic, and I switch between face_yolov8n and face_yolov8m for face detection, hope this help =)

@ColorWolve thanks I'll try this later today

for some reason the adetailer was messing up male faces for me and making men look like women, with mascara and eyeliner and such lol

@makaveli_313

Oh, that's very complicated to solve. I have a similar problem, so I can't help much, but I can tell you that using [SEP] can make each character's face look different. on adetailer

For example: if you have two character on the photo

a girl with red lips[SEP] (for the first character

a boy with a beard[SEP] (for the second character

The [SEP] token will separate each character, allowing them to have different facial features

@coyotte  You’re welcome! by the way Thanks to your model, and I hope you enjoy the results!

@ColorWolve Agreed!!

@coyotte Is it something that can be done on model end for the eventual 5.0, or difficult to fix

I noticed the same, anyone has a fix for that?

find an example image of what you're looking for and try something like this https://huggingface.co/spaces/fancyfeast/joy-caption-alpha-two i use that with a long description and usually get the angle I am looking for

@Plinkerton thank you I'll try this : )

How?? 👀

@Pebiss AI influencers - growing their socials and selling NSFW pics.

@aileverager555 That actually work?? I just figured that was a meme

@Pebiss i thought it was a meme too!! XD i'm flabberghasted someone has managed to make money with this (generative AI art).

@Pebiss It does work. Of course. It's basically an artificial OnlyFans girl. Huge money here

@coyotte There are real celeb on Pony, with real name. What are you afraid?

Civitai purge just the most famous influencers, whos lawyer just send a nice letter.

You can fuck real celebs in Pony, in other SDXL models with loras, na meh.

@kyvonzorawar999 There's a million ways to bring face consistency back into any model. You can use FaceID, InstantID, Reactor and tons of other ways. Stop asking dumb noob questions and let the model creator focus on fine-tuning the NSFW concepts. Everyone expects a single model to do everything but its better to have multiple models specialized to do 1 thing excellent and bring them together in a ComfyUI pipeline

@coyotte We have one of the best creator, who is a crybaby. Fucking shame.

@damiang And every SDXL model have built in celeb likeness (prompt a name), no need for third party extensions for everything. He purposefully removed that.

@damiang And yeah, you would cry, if you knew, what Pony is capable in realism with good prompting.

And Pony has detailed expressions with just prompt. SDXL has two emotions: smile and emotionless.

SDXL is for noobs.

@coyotte Yeah, you are like my one of the fav indie game developer, whos game was very unique and high quality, but lacked of a ton of QoL, and he purposfully didnt hear anyone about to add something, which makes his game much better.

@coyotte OH, sorry

@archerollivander127 If Pony can do realism, has the versatility for being Pony and has celeb content. Then why are you CRYING over this model?

But hey put your actions where your mouth is. Send me one good realism prompt example for Pony and let's see how much full of shit you are

@archerollivander127 If you love pony so much, then why come here for the SDXL model and cry about it? Go back to pony and be happy and make all your celeb pics. No need to come here and cry like a baby.

@damiang pony is terrible at realism. It just can't do it.

@ElectricDreams Yes I know it can't do it. I was talking to that gooner that was crying about some decisions made by @coyotte. The best way to use both worlds is generate 4-6 steps in Damn Pony and then push that latent image into Lustify and do additional 4 steps to bring back the realism. This way you have the prompt adherence and versatility of a Pony model but the realism of Lustify. Although I find Lustify great on its own already. But this workflow of pre-running steps in your fav model and then finally baptize it into Lustify works exceptionally great. If you have Lightning models it's blazing fast

@coyotte Yes exactly, my comment was inspired by your workflow. Should've sent the link. Thanks for clarifying

@coyotte That's what it said under the recommended parameters, no?

@coyotte With denoise set to 1 it works fine. Thank you. Misread the instructions :-(

@coyotte So V5 will be the final for SDXLyou mean to say, or the last NSFW in general from your end.
won't bother with sd3.5? or flux etc.?

@coyotte can I try RC? :3

Guys, let's tell it like it is.

Lustify should simply be considered, as it was for Pony, a new starting point, a new Base model from which fine-tuning can be done.

Civitai should take the model that this man made and add the "Lustify" section, people should create LoRa and do training exactly like they do with pony, like they do with SD1.5, with Flux etc...

If and when the "final" version comes out, it should not be called 5.0 but rather simply "Lustify" and from then on it should be used by everyone.

I swear to you @coyotte , I'm a poor shitty man without a penny, but if I had the opportunity I would gladly finance your work in order to help you as much as possible and others should do it too, simply because it's the right thing to do.

Oh, and by the way I was serious about the fact that your model should really have its own section and instead I get to see other crappy models that I can barely follow the instructions given through the prompts... 🤷‍♂️

@BobDF loved your comment : )

@coyotte fair enough. Great job on your effort, you made the best nsfw model ever, as you can tell by others feedback as well : )

BigAsp2 dataset is 6.5 million, and doesnt have its own section. Lustify maybe several thousands new data.

@bronsencaston766 Well, since BigAsp is clearly inferior (and I'm not saying this, just look at the data and user feedback), I don't understand the comparison...

@BobDF my results with bigasp 1 and 2 are superbad, like garbage-tier monstrosities. Maybe i don't know how to use it. It is only me? It's a checkpoint it confuses me. I don't know, maybe i'm the only one who has those problems.

@ElectricDreams I know it's a very strange model, it needs very particular settings to generate correctly, and as far as I'm concerned it's another big point against it.

With lustify, even with a simple sentence and without negatives I can bring out superb generations, so as I already said, they are on two completely different levels.

@BobDF Of course easier to prompt Lustify. Its just a SDXL finetune for porn. Bigasp is like Pony, with own language. But we have no dictionary for it, while there is for Pony. Thatswhy its shit in this state.

I reflected, who said Lustify needs its own section.

@austinbrandtly992 ok I'm glad to confirm i'm really not that bad. Yeah it's a very difficult checkpoint to use. Not exactly very friendly.

@ElectricDreams yeah, it really needs more finetunes, merges.

@ElectricDreams will you plan to post the merge with cyberrealistic?

@coyotte They considered that the girl I shared was too young (too cute I guess), so I suppose you didn't see her. I'll share another one that is still very cute, but looks less young, I hope. I'll have to experiment much more to make what I really want!

@coyotte have you tried the DPO lora? in theory its a lora for training checkpoints that expands/enhance prompt adherence. https://civitai.com/models/242825/dpo-direct-preference-optimization-lora-for-xl-and-15-openrail?modelVersionId=273996

@coyotte No? 🤔 I still see it. Go to my profile, the only (now two) pictures that are not "drawings" are made with your model.

@coyotte Oh, really? Damn it. After they deleted the first one it seems like I wrongly put them in the wrong category! And someone has reported them... again... Must every girl look like she is in their 30s or what?!

Where do you put the DMD2? in a unet folder? and use a diffusion model?

@coyotte Could you recommend a specific Upscaler? and Hires CFG scale for the same please

Update me on that