Update 10.07.2026. Lustify V10 (Krea 2) IS OUT!
Recommended workflow.
Alternative Access & Support: Outside of CivitAI's Early Access, you can get the model via my Boosty.
As always, after the early access ends, there will be a 2-week period where the model is available for "on-site generation only." After that, it will become available for everyone to download. The open release date: 10/08/2026 (10th of August).
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
FAQ
Comments (87)
Hey guys, V9 (Zenith) is out! This is going to be the last SDXL version. V10 will be on a completely different architecture (I'm working on 2 different ones right now, and whichever works out best will be released). I'm pretty happy with what I've managed to achieve with this architecture as a single person without a professional rig. Thank you so much to everyone who has stuck with me so far, and I hope to see you again when the new architecture version releases! Enjoy V9.
I've tried my best to improve the lora training on this model. It should be ~10-15% easier now. Not a lot, but if you always felt like your character loras had missed this tiny notch of resemblence - this might be big for YOU.
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
I love you @coyotte . I can't wait to see how your magic works on newer models.
Great work coyotte! Any idea on what architecture you are going to base the next model?
Bro we stopped using sdxl a year ago
Define "we". I still do, even though I can run klein 9b/z-image/chroma just fine. Besides, this is not the "one-year-old sdxl", as I've been training it on 1536 res for V8 and V9.
a lot still use sd1.5, forget sdxl. old models getting support is good
thanks for taking the time to make such a pointless comment !
@goeypants400 But why?
@mariedoll123 No problem
I stopped using SDXL, except for when coyotte puts out a Lustify model. They're so good that I can't care that I have to go to SDXL. I have a feeling that if the same training went to Z-Image or Klein 9B, that you wouldn't get nearly as good anatomy. The best results I've gotten are using Lustify, then doing a pass in ZIT to upgrade skin details. The best of all worlds.
@MonkeyForever its preference and fun. liking the certain look from it, its specific quirks, and other opinions. and because generating itself is fun, than just always using the same latest model
Nah. I still prefer SDXL over all the other new models that requires NASA computer to run. Speak for yourself.
same guy who suggest to use sdxl model for great realism with other models on his discord btw , lol
@markhassain3712 I really dont do that at all actually
@Jellai you mind sharing yur workflow?
@Wuastau Here you go. The link will work for a week. This isn't the exact workflow I use, as I tend to cobble something together on the fly with whatever I'm doing, but it's the one that some impressive images were using.
https://limewire.com/d/4D2C6#e7NZEkhPJP
And here's the post with the images:
https://www.reddit.com/r/unstable_diffusion/comments/1to41k1/morning_walk_in_the_park/
Oh wow, just when I thought that v8 can't get any better, you release something even more polished. You can't keep getting away with this 🤭 Jk, that's sad that you don't plan to continue the SDXL line of Lustify, but that's understandable. Looking forward to the new stuff. Any hints on the architecture?
P.S. The man x monster girl and girl x monster man additions were really needed. Thank you!
u going for ZIT now?
I have a ZIB and Boogu in training right now. Whichever works out best will be released. If Krea 2 comes out with a good licensing, I might switch to it.
@coyotte yeeee nice
@coyotte how about Qwen 2512? It’s better than zit imo.
@coyotte Check out Ideogram as well. Pretty neat model.
8k for that just lol, git missing fingers all around your samples, and it does not look like an sdxl at all, did you mixed some illu or pony in it ?
"Missing fingers all around your samples" - there is literally one image where a finger is covered by a thumb. Saying it's "all around" is a massive stretch.
"does not look like an sdxl at all, did you mix some illu or pony in it" - First of all, both Pony and Illustrious are based on the SDXL architecture, so saying it "doesn't look like SDXL" while suggesting them makes zero sense. Second, you might have missed the "checkpoint trained" label. I've created synthetic data to train the monster interactions (for obvious reasons, since they don't exist in real life), but there are no merges here. And no, I did not use pony or illustrious to create that synthetic part of the dataset, but I did use Chroma for that.
"8k for that just lol" - No one is forcing you to be here. If the model is not for you, you can simply close the tab and move on. 8k Buzz translates to roughly $4. Meanwhile, I've spent around $530 out of my own pocket on GPU rentals for this specific version alone (and that's actually on the cheap side for a proper full finetune). But I guess as a person who only creates merges, you wouldn't know anything about the actual costs of training.
Dude you create 2.5D anime slop, do you seriously think you have a say there?
@mskiller51 Didn't expect to see a creator of your scale disrespecting other creators. That's a massive -rep for you. Envy is a bad thing 🥳
@coyotte Based.
@coyotte has consistently put together masterful checkpoints, pushing SDXL's limits, with great capabilities, and achieving a unique look, which is really hard to do at this point. @mskiller51 , you are making yourself look like a fool with this comment.
@letunchik if you think so, you might have not checked all my work but that's ok, beside I do prefer anime to realism.
@lunaxo sorry to have been disrespectful, I just gave my point blank opinion, maybe too much straight forward, but you did way better in the past without asking for that much. and your samples does all miss fingers, that's a fact.
@Jellai regarding masterful and pushing sdxl boundaries, if you only use sdxl I would say yes, but no, I am sure this is not only sdxl, and I did enough model mixing and testing to know. like said to lunaxoi, sorry if I sound too straight forward, i was just suprised! @coyotte does indeed better work usually, just this last one, according to his samples, well ...
@coyotte did not reply to you, sorry, but you are right, pony and illustrious even noobai are sdxl, I am fully aware of that, but either you overtrain or something, your samples, when I look at them where not good, and now it is no more the same samples as it were when I post my comment, indeed, your new samples looks way better.
I am aware of trainning, there is training lora and mergin them in models and real model finetuning, I know you do excellent work in past, I was just disappointed to see both those samples and that price tag, but you change those, they are now better showing your model result, and your cost explannation, indeed trainning is a cost to everyone, I did not see illustrious, pony or other huge model finetuning asking for that, but hey every one has their own views and needs.
And sure you got a huge community that did not like my comment but hey, I just tell my thought
@letunchik says the one who never did nothing :D
@mskiller51 I did not change the samples. At all. These are the exact same images that were there when you posted your comment.
The reason you don't see other big models asking for $4 for the early access is because the vast majority of models out there are just merges, even if they claim otherwise. As for the actual full finetunes, they are usually made by large groups or teams with corporate sponsors, or some rich furry guys who have more money than they know what to do with xD I don't have any of that, though I wish I did
@coyotte the big models I spoke off were kinda pony xl 1.0, illustrious xl 1.0 or even juggernaut who I believe are all trained, but yes, like said, everyone sees his way to move forward. And sure I just merge, so I don't ask for anything as there is no real cost to it.
@mskiller51 Dude, Pony, Illustrious, and Juggernaut were all heavily sponsored. The creators of Pony and Illustrious had massive financial backing/GPU grants to train those architectures, and Juggernaut is literally sponsored by RunDiffusion.
That’s exactly my point - you are comparing a solo independent creator paying for GPU rentals out of their own pocket to projects backed by corporate sponsors and platforms. It’s a completely different scale.
@coyotte indeed, you are right, my bad.
@mskiller51 Your bad on the entirety of the posts you've made here. Like you basically got everything wrong. You doing ok? This is a weird crash-out.
@StanleyPain you do not feel a little late on your comment lol I belive I clarified myself, but if you feel the need to lash out more, go ahead§§§ :p
@mskiller51 Somewhat indicative of your state of mind that you think my comment is "lashing out"....I'd seriously consider a bit of self reflection here.
@StanleyPain thank you soo much for your concern ! 🤣🤣🤣
Hello @coyotte . You are amazing . I've tried v7 , v8 and i didn't get anywhere else such a good and detailed results for nsfw . but i had problems with the trainings ( never get a good likeness ) . I am not an expert , i do not know much about trainings and workflows. I see that you say LoHA + Prodigy + Cosine. 16 rank, 16 alpha / 8 rank, 8 alpha ( i din't test it yet ) . Can you please give me a full setiings config I've used Ai toolkit 20+ trainings different settings with differencial guidance everything ? Thank you for all your work !
Any plans on DAMN v6? This is still the best realistic checkpoint for illustrious imo and seeing lustify getting updates while damn is left behind feels disapointing
How does the model work with Character Loras?
anyone ?
@contactmayabanu734 if you want to use it with character loras trained directly on the base SDXL - it will work, but it won't be perfect
@coyotte i want to train on v9
Any uptade on your z image models?
Z-Image was turning out pretty meh (nothing you couldn't do with the existing loras), the penises looked bad and vaginas looked used up. The same problem everyone has finetuning the Z-Image basically.
I was about to start the Boogu train, but Krea 2 droped, and now I'm pretty sure I'm pivoting to it. It's has good quality, good style understanding, decent licensing, Turbo and Base models. Though I might proceed with Boogu training in parallel, or Anima might take it's place.
@coyotte oh maan what a shame,I guess you will skip Klein 9b and ideogram as well, please in any case focus on boogu first please
@Neon_signs any particular reason you want boogu first? In most cases it's way worse than krea 2.
To be honest, Z-Image needs to go. It's fundamentaly broken as it doesn't work with more than 2 loras (in the best case scenario, in the worst case even 1 lora can be the limit).
Klein 9b has a bad license, and Ideogram is tiresome to use, as you need to have an idea of what exactly your composition should look like to fill in all the json data. Besides, ideo's licensing is even worse than Klein 9b.
@coyotte well I thought it has some potencial in the realism part , but that didn't happen so well and has some weird grid lines like qwen, but now I will scratch that honestly and straight out go for krea lol, even tho needs a real fine-tune for reality, the amount of characters that knows is insane ,so just keep focus on krea and anima please,can't wait to test your stuff in them
Krea-2 is the way to go, but please consider releasing both the raw (base) and Turbo versions. The Turbo is fast, right, but it's very dense and stubborn the way Z-Image Turbo is: there is virtually no variety between swipes. The raw version is way more creative, and it takes the negative prompt, too.
V9 is a big improvement over V8. Thanks for the hard work.
Glad you're enjoying it
What would you need to train on aesthetic? what dataset and captioning?
I write this, because i spoiled myself with the grok image model(quality mode). This means, a model can be photo(not just photorealistic) and beautiful together, which is a problem in local models.
Do it on Krea2 please :o
Already in work, and it will come out much sooner than you think!
@coyotte Hahaaa you're the goat
@coyotte please please include support for massively obese men when you build the krea 2 version. please please please.
@Grandpa i'll see what can be done. It's already in training, so I might not be able to add anything else into it right now
why I can't download
Please read the big red bold text in the description, that was there for 2 weeks.
I managed to download it when it went out of paid access.
For some reason I'm getting much lower quality outputs than in previous versions; using the settings you listed in the OP. Is there an example workflow for V9 by any chance? Thanks!
I agree, I don't like version nine at all compared to Apex.
I'm having no problems with v9 whatsoever, IMO it's way better than v8
i'm getting worse results as well, but that's been true of every new release, they all have some special (always undocumented lol) quirk and you just have to figure out what it is by trial and error.
yoooooo coyotte i didn't realize I wasn't going to be able to download it, i waited out the entire week chomping at the bit, now what the heck do i do
edit: nevermind, I finally saw the red text in the description! i can survive another week. xD
Hello, I saw your reply mentioning that you're already training the Krea 2 model. I also noticed earlier that you planned to train a realistic Anima model. However, Krea 2 is quite demanding for my 8GB VRAM graphics card and difficult to run properly. I hope you can release the realistic Anima model. Thank you!
Anima one is in work! As for Krea - I'm pretty sure you'll be able to use Q4 quant version of it just fine
It's RAM+VRAM, so unless you're already specifying (in which case RIP), you could likely run fp8 and just have to wait (if you cant run newest cuda)
Hey coyotte, love your work! Do you have a community anywhere (even a paid one) where we can ask questions and follow your progress? Would be awesome to keep up with the Krea 2 and Anima checkpoints. Thanks for everything! 🙏
Thank you so much for the kind words!
As for the community - the sub for my boosty grants you an access to my Discord channel. But you can ask me questions there in the comments and I will answer if it's possible.
Although the discord channel has some news on training progress with the previews.
Weak v2.0 inpaint,Not n*s*f*w apex inpaint
Honestly you'll be better off using Flux Klein 9b with SubtleShader's finetune or SNOFS lora. SDXL inpainting models are very outdated and limited.
I only made the apex version for it because people asked me to.
I cant' download....
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.
-> should be available for download in the next few days
How much did it cost you to train Z-Image, and how much is Krea 2 costing you so far? Once Krea 2 is ready, are you planning to release the Base, Turbo, and FP8 versions? Are you also planning to create LoRAs for large models like Qwen or maybe Chroma? And do you have any plans to collaborate with other creators, like SubtleShader, iamddtla, or Danrisi, to combine datasets or knowledge and eventually release a model that has everything? Approximately how many images do you usually use to train your models? One last thing: do you have your own community, or are you very active in any Discord or Reddit community?
1) As for the Z-Image - it's irrelevant, it's impossible to make a good finetune with good-looking genitals, so I dropped it. I don't know what they did to the base model, but it's just the way it is. The lack of good NSFW finetunes proves it.
Krea is burning a lot of money as I have to rent H100 NVL (and it's barely enough for the finetuning, insane), but it's very promising. It's 3$/hour for the last 7 days, and there's plenty ahead.
2) Yes, Base + Turbo + FP8.
3) Not, not really. Me and Subtle have pretty different views on a lot of things, it wouldn't work. And as for Danrisi - I've never really talked to them.
4) The current Krea train uses 15k images. I could've used 3x more (almost the whole dataset), but then it would cost me 3x more as well. I'm pretty sure with heavy architectures like that it makes more sense to add stuff in increments, build on top of what you've built earlier.
5) I only have my own Discord channel for boosty subs. There's some exclusive stuff (mostly training progress, previews before release, etc).
@coyotte Krea was apparently trained on a very superior dataset compared to something like ernie or other newer models in this weight class since it has a way higher understanding of difficult anatomy and real-realism. Do you think a 15k images (mostly nsfw bias i guess) full finetune will satisfy without overcook certain params into the whole thing? I really hope you can work some magic here but imo for the customer grade hardware community only a 6b-ish flowmatch model pretrained on high class sh*t free of cencoring and running with 4b qwen text can become a replacement for sdxl
@ax_hgs I think you're underestimating the variance 15k images can cover. Besides, I don't plan to limit the training to a single 15k batch; the plan is to make it a tiered process. I'm starting with an initial 15k, followed by subsequent batches of another 15k, then more, with further adjustments determined by the intermediate results.
Regarding the architecture: we don't have a viable 6B-range model with a permissive license. Z-Image is the only real contender, but it has fundamental flaws (like incompatibilty with more than 2 LoRAs, or significant difficulties to make the genitals look 'okay' at least) that make it not worth the overhead. Klein 4B is just plain awful.
@coyotte Isnt training with multiple runs rewriting the model plus the previous run at some point and requires a even lower LR? Do you tried promting mid-run for a actress plus a complex pose not covered in your dataset but both known by krea to check if things start to drift? May i also ask why licence is a thing? It seems noone rly gives a fk about it and klein9b is still most promising for its size.
@ax_hgs Not necessarily.
Handholding wouldn't really rewrite the doggystyle from previous run that much, just like lighting styles wouldn't rewrite particular lenses. As long as the data in each run is sufficiently distinct, it won't drift. Besides, you can also use the previous dataset with a very low repeat amount (e.g., 0.2) as regularization to ensure core capabilities are preserved.
Obviously, training alll in one go is the best approach, but I can't afford $3k per version. Early Access doesn't cover that kind of overhead, and unfortunately I'm not some rich guy that has more money than he needs.
As for the licensing - I do give a fuck. If the license prohibits commercial use, I can't do Early Access. And if I can't do Early Access, I can't train the model, because, again, unfortunately I'm not some rich guy.
Anyways, I've shared some of my progress in my discord channel - I think it's looking pretty good. First Krea version will be out this week I think.
@coyotte
> but [Z] has fundamental flaws (like incompatibility with more than 2 LoRAs, or significant difficulties to make the genitals look 'okay' at least)
So glad to see people finally say this. It's been UPSIDE DOWN WORLD watching people hype up Z for months and months. That trash was released broken and never got better. It makes a good portrait of a girl, that's all.
Glad it's dead. Glad people like you are on to Krea.
I know people are salty that it's GenOnly. But you can purchase it from the Author as well. Support good developers! They gotta eat too!