Update 19.06.2026. Lustify V9 (Zenith) IS OUT!
This is the absolute final version of Lustify built on the SDXL architecture. You can find the full changelog and a breakdown of the new features there. <----------
Recommended Workflow.
Alternative Access & Support: Outside of CivitAI's Early Access, you can get the model via my Boosty.
Just like last time, after the early access ends, there will be a 1-week period where the model is available for "on-site generation only." After that, it will become available for everyone to download.
Alternatively, you can make a direct Bitcoin (BTC) donation, just make sure to DM me after that: 1CoMZvrXtRXbz7BCzdvxifSjHqhaf1yUZs
LUSTIFY
Lustify is a photoreal checkpoint optimized for creating high-quality imagery of women in explicit/sexual scenarios. The model has been trained on a hybrid dataset, meaning it natively understands both Danbooru tags and natural language prompting.
Note: While it has a strong NSFW bias, if you guide it properly, it is fully capable of generating stunning SFW photography, objects, animals, and fantasy environments.
Crucial Points Before You Download:
NOT Pony or Illustrious-Based: This is a standard SDXL checkpoint. Do not use Pony/Illustrious LoRAs with it; they will not work correctly.
Distant Shots & Anatomy Limitations: Like 99% of SDXL models, distant shots can sometimes suffer from distorted faces or hands. This is an inherent limitation of the base technology (4 channel VAE), not the finetune. It is incredibly easy to fix using Inpainting or ADetailer. Please utilize these tools rather than reporting this as a bug.
Experimental Nature: My training workflow involves a lot of unconventional, high-risk experimentation. Because of this, some prompt combinations that worked perfectly on other checkpoints might behave differently here. If you encounter unexpected behavior, feel free to give feedback so I can look into it.
Recomended parameters:
Sampler: DPM++ 2M SDE/DPM++ 3M SDE,
Scheduler: Exponential/Karras
Steps: 30
Cfg: 2.5-4.5.
Highres.fix: upscale by 1.4-1.5, denoising ~0.4
For the distant shots of people you probably wanna use highres.fix and/or adetailer.
Also, use reasonable positive/negative prompts, shizoprompting does more harm than good (with any SDXL model).
Some tags that have a big visual impact with this model:
Camera type tags: "shot on Polaroid SX-70", "shot on Kodak Funsaver", "shot on GoPro Hero", "shot on Canon EOS 5D", "shot on Leica T".
Photography styles: "analog photo", "glamour photography", "street fashion photography", "candid photo", "amateur photo".
Lighting types: "cinematic lighting", "neon lighting", "soft lighting", "dramatic lighting", "low key lighting", "bright flash photography", "warm golden hour lighting", "radiant god rays".
Film types: "Ilford HP5 Plus", "Lomochrome color film", "Fujicolor Pro".
Photographers: Alessio Albi, Martin Schoeller, Miles Aldridge, Oleg Oprisco, Tim Walker.
Others: "film grain", "bokeh", "dreamy haze", "technicolor", "underexposed", "low quality", "lowres"
Alternatively, you can give a read to this wonderful prompting guide by PromptGeek. It has A LOT of useful tips I didn't know about even after years of using SD.
Description
Lightning version.
Recomended parameters:
DPM++ SDE
Karras / Exponential scheduler
4 Steps
Cfg: 1-2.
Highres.fix: upscale by 1.4, 2 steps, denoising ~0.4, upscale model "nmkd siax" or "nmkd superscale"
Due the nature of "lightning" models, this version might follow your prompts worse then the regular one. Also, I've noticed that the lighting type tags need more weight in this version, for example "(neon lighting:1.3), (cinematic lighting:1.1)".
FAQ
Comments (78)
Great model,Very strong movie feeling, recommended
Thank you!
Here before it blows up! :D
when? 👀
Hello first at all thx for the model.
Is there a special reason why you merged it with pyros stuff?
2. If you merge a model with a pony model provide more information pls.about the complete score_ tag for example know what I mean.
Like you did it with your info about,cameras,light etc.
Ok lets play with your model now :-)
Hello!
1) The reason I've used pyro's nsfw checkpoint in the recipe is because it's just a good porn checkpoint with a lot of concepts. It's about only 15-20% there tho, as it was replaced by the weights of my personal finetune.
2) No, none of the pony models were used for this model. bigASP uses the "score_" tags, so that's where this is coming from. Also, the dataset that I used for finetuning that mix is from my other checkpoint "DAMN!", which was a pony model indeed. I didn't change tags at all, so some of the pony/danbooru syntax is now present in this model.
@coyotte Hey thx for your quick answer bro :-)
About score I have never tried to weight the score_ system it sounds interesting. Can you gimme a quick hint how to do this the best way please?
@mscorpion7285 just use "score_9" tag in the beginning of your prompt, and that's it. "score_9" tries to give the best result in terms of aesthetics. "score_8" and "score_7" will be a bit more amateurish (although with this model this won't be as noticable as with bigASP".
By the way I found your pony model,seems to be interesting and a worthy download. thx for that model 2
@mscorpion7285 i'll be honest, my pony model is deprecated at this points. i might update it some time in the future
@coyotte okay.thx for your honest opinion.
@coyotte damnponyv2 is one of the GOAT checkpoints. Don't sell yourself short bb.
It's the only photoreal model that understands the male character LoRAs that I throw at it.
I'm a simple man of culture, I see a finetune of BigAsp, I pound the download button like a redheaded stepchild.
LMAO XD
bigASP is such a great checkpoint for further training, @nutbutter did a terrific job on it
@coyotte I'm knee deep in training loras against bigasp. still nailing down training rates, but i ran a test on a lora that finished overnight, and while a little overcooked, it works well on your model. be a few days before i've got one "just right" but once i've got the right balance, the datasets are all ready to begin pumping them out.
you pound redheaded stepchildren?
@shapeshifter83 What are you doing stepbro?
I like how the skin details really! pop! on this model. Plus it really nails the sky high platform heels that I have a fetish for.
EDIT: Ohhhhh yea! baby!, using (sharp focus:1.2) really made the images much sharper. All that is left now is to merge it with my private model and bam, perfection. EDIT2: Just merged it with my private NSFW model and the results are amazing.
TL;DR- This is definitely one of the best 'REAL' NSFW models that has come out in the past 5-6 months. And trust me, Ive tried every single one because Im a degenerate and always seeking the best. I dont do pony, its trash for realism and mixing with realistic models ruins the realism.
Thanks a lot for the kind words, I'm glad you are enjoying this model!
How does one get access to your amazing merged checkpoint Fox? :D
What does your private model handle which this one does not?
@udf2wozza I wont release it due to being on a data capped internet via phone hotspot, and slow connection disconnects causing repeated failure attempts using Civits upload system.
Also the fact that its just a mix of various NSFW models found on this website anyone can do it. I just managed to find a recipe that works for me.
@countlippe nothing much, its just a different aesthetic that I prefer, its all subjective.
Can I use it as inpainting model?
You can inpaint with this model, it will work, if that's what you mean.
But it isn't an inpainting model in it's nature. Look into brushnet or fooocus to use literally every model as if it was an inpainting model. Including this one
thanks!
@coyotte I second fooocus. It works great at inpainting with all the models I've tried.
Did anyone manage to use this with Fooocus?
I'm not sure what I'm doing wrong, but the pictures come out looking really weird, like oversaturated cartoon with strong outlines. Even when using the same prompts as in the examples here.
The lightning model works better, but the result is still very... kind of grainy / over sharpened.
Thanks! Sorry if this seems like a dumb question.
One moment, I'll install focus to see what's wrong
Okay, I've figured it out. Turn off "Fooocus Enhance" style
@coyotte Oh indeed! Thank you very much! Lightning is still weird, even after removing all the styles, but it did fix the issue for the main non lightning model.
@Larxian are you using the recomended parameters for the lightning? because it might behave weird on other samplers except the ones i've listed in the "about this version" tab
Recomended parameters:
DPM++ SDE
Karras / Exponential scheduler
4 Steps
Cfg: 1-2.
Highres.fix: upscale by 1.4, 2 steps, denoising ~0.4, upscale model "nmkd siax" or "nmkd superscale"
@coyotte I see, I did see the recommended parameters, but I don't think they can be used with Lightning in Fooocus, because all these advanced parameters are locked when you switch to Lightning, unless I missed something.
I just checked my output and yeah, the pictures rendered in lightning are using euler sampler and sgm_uniform scheduler, so that answers it.
Thank you anyway for your really fast replies, I guess I'll just use the non lighting version, this model looks promising!
Edit: Actually, maybe it's because of the "performance" mode in fooocus, I thought I had to change that to lightning too, but it seems like using speed and manually changing the steps to 4, and the other settings, while using the lightning model, works better! So it's on me, sorry for that.
@Larxian I was about to write about that part with modes, but you've figured it out yourself :D
Yeah, u're not supposed to check "lightning" mode when you load a lightning model. It's for regular models only. This mode applies a lightning lora to a checkpoint, essentially making it a lightning checkpoint.
So what you can do is to load the regular one and choose "lightning" mode. It will give about the same results, as the checkpoint I've uploaded (but not 100%, because I've made some tweaks to it).
Or, just use it as a regular SDXL checkpoint. This option has it's own strengths
Its most likely the VAE, Foocus does not allow you to use a custom VAE, its locked to the standard OEM VAE.
@fox23vang226 that was a long time ago. now you can set any VAE you like
Do you recommend CLIP skip 1 or 2 for this model? Just starting to tinker with it and seemed to work okay at 2, will try 1 also but since it has some Pony in it that's why I ask.
It doesn't have pony in it. It's just the dataset I've used for training was tagged as if it was for pony. You can find the insight into how this model was created in the end of the description.
If you're using comfyui - leave clip skip at - 2 or, better yet, don't touch it at all. Like, remove the node that changes clip skip.
If you're using automatic/forge - doesn't matter, clip skip doesn't do anything for sdxl models in it.
@coyotte thank you. I was mistaken because of the score_9 style prompting and assumed it had some Pony in it. I have mostly been using Pony models lately so didn't realize that CLIP skip doesn't matter in my case (I use Automatic1111 at present).
i've gotten so adept and in-tune with natural language prompting that i feel like i can't even do danbooru/tag style prompting anymore :(
This model can understand natural language, so that's fine
Also there's an extension for you: https://github.com/DominikDoom/a1111-sd-webui-tagcomplete.git
I can't live without it. It tells you what the best booru tag is for what you're typing
This model is the reason I finally decided to switch from 1.5 to give Pony a try and damn, what a model! The anatomy is fantastic, from hands to feet shots. Seriously, thank you for the work you put into this, following it to make sure I catch any updates for this model if that's planned ;)
Well done!
Thank you for your kind works!
I have to tell you - this isn't a pony model, just a regulard SDXL based model. I encourage you to try out some Pony models tho, they're completely different type of beast. I'm gonna update my Pony model "DAMN!" pretty soon ;)
@coyotte Oh I got confused after reading the description, the word Pony stuck to me haha, I meant SDXL. You've earned a follow with this model, it's so easy to prompt for different positions and the skin details are fantastic, can't wait for V2!
where can you download sd1.5 from
HUGE fan of this model. Image composition/lighting/details all look amazing. Skin details in particular are great. Sincerely hope you keep working on it. Feels like we're getting closer to pony levels of image flexibility. The only downside is I'm having the hardest time generating vaginal penetration, which seems pretty important lol. It generates anal 9/10 times, even with anal/anal sex/anal penetration heavily weighted in the negative prompt.
Funny thing is that I have a completely opposite problem xD
Don't worry, this is a V1 version just to gather feedback and see what people want to generate. The really tasty stuff will be in the V2
Wow this is a real good model, love it
Really nice job, only issue is that my female model lora's loose a lot of likeness when using with this checkpoint
It is very far from SDXL-base weights because a lot of finetuning was done. This inevitable and I won't compromise quality for compatibility. If you really like the model, you can train a version of your lora for this specific checkpoint.
@coyotte yep I did training now on your checkpoint directly, results are way better! Cannot wait to see the next version
Any chance of posting the unreleased finetune?
What unreleased finetune? Where'd you see that lol.
@welcometoea in the description, probably. The very end of if.
And no. I didn't release it for a reason. It isn't as good, has serious aesthetic problems and generally not convincing enough. I'd prefer to leave that part of "lustify" behind the scenes.
This is the best NSFW SDXL model created to date. it's like Pony but it's not, you won't get the same rounded face with big forehead and eyes, you can get different more realistic faces instead.
A big thumbs up.
The best Lightning model out there! Hands are quite tricky in some cases, but I figured out a good workaround. I am very excited for the Versions to come! Chapeau!
what settings did you use? i can't seem to get the lightning version to actually accelerate with typical Lightning params, i had to use the regular model and a Hyper LoRa for any noticeable speed boost.. is this a slow model in general?
@Narz I use the App "Draw Things" on a Apple MacBook Pro M1 Max. It is very fast, 15 seconds per generation
@Narz are you using the parameters I've recomended using? "with typical Lightning params" seems like you aren't.
@Narz
For lightning checkpoint, try these parameters:
DPM++ SDE
Karras / Exponential scheduler
4 Steps
Cfg: 1-2.
Highres.fix: upscale by 1.4, 2 steps, denoising ~0.4, upscale model "nmkd siax" or "nmkd superscale"
what CLIP skip should i use?
If you're using comfyui - leave clip skip at - 2 or, better yet, don't touch it at all. Like, remove the node that changes clip skip.
If you're using automatic/forge - doesn't matter, clip skip doesn't do anything for sdxl models in it.
@coyotte wait, wtf, really? auto1111 ignores clip skip? have I been hamstrung this whole time?
@volatilevisage2 yes. https://files.catbox.moe/ivmtml.png
But that's the case only with sdxl/pony models. It works fine on 1.5-based checkpoints
Tip: Use a flexible semi-realistic pony models as the base as they typically give good prompt adherance, then use this as a refiner to get photorealistic results
Great tip and that's what I've also figured out.
My updated pony model is coming out soon (like today, if all goes well), so it can be a great opportunity to try it out 😌
what do you switch at when using a refiner?
@goujons180 Played around with different values, but typically about halfway (0.5), or does highres fix using this model.
@coyotte I AM SO HYPED.
Should i use sdxlvae?
Not necessary - I've baked VAE into the checkpoint.
@coyotte cool, thanks, and for ComfyUI as well?
I really wanted this model in onnx format
I think you should look into the scripts for converting models.
I mean, I could spend some time to figure out how to do that and convert it for you, but you're better off doing that yourself. That way u'd be able to convert literally any model you like later.
https://www.youtube.com/watch?v=hE-dSzVSIbI
Lustify V2 this week, hopefully.
Training in progress, already made a couple of training runs for 15 epochs and still trying new things to get the best result. https://files.catbox.moe/zzitan.png
It is possilbe to add futanari or it will break the model too much? Thanks for this amazing work.
I don't think I will break the model, but the problem is that V2 is already baked and waiting for the upload. I'll add futa at V3