Alternative Access & Support: Outside of CivitAI's Early Access, you can get the model via my Boosty.
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
Available Model Variants:
Raw:
fp16/fp8/int8_convrotTurbo:
fp16/fp8/int8_convrot/Q4 GGUF/Q2 GGUF (my custom implementation)
Hardware Recommendations:
4 GB VRAM: Either Q2 or Q4 GGUF variant.
8 GB VRAM: Your best bet is the Turbo Q4 GGUF variant.
12 GB VRAM (or if you don't mind slightly longer generation times): Go with either Turbo int8_convrot (after the latest Comfy optimizations this is the preferred variant) or Turbo fp8.
24 GB VRAM: You can comfortably run the full Turbo fp16 variant.
Use Turbo for everyday inference (generation) and Raw if you plan on training or fine-tuning.
FAQ
Comments (73)
Krea 2 is out.
Available Model Variants (find these in the 'Optional Files'):
Raw: fp16 / fp8 / int8_convrot
Turbo: fp16 / fp8 / int8_convrot / Q4 GGUF
Hardware Recommendations:
8 GB VRAM: Your best bet is the Turbo Q4 GGUF variant.
12 GB VRAM (or if you're 8 GB VRAM user that doesn't mind slightly longer generation times): Go with either Turbo fp8 or Turbo int8_convrot.
24 GB VRAM: You can comfortably run the full Turbo fp16 variant.
Use Turbo for everyday inference (generation) and Raw if you plan on training or fine-tuning.
omg everyone get in here (and then don't make eye contact u weirdo)
Aw hell yeah dude. Just in time for the weekend
@coyotte After working with SDXL for so long, how do you feel testing the waters with Krea 2? It looks from here like it picked up training REAL well. Coherent naughty parts out the gate. Was Krea 2 a pain to work with? Do you think it picked up the training well?
@Jellai Krea 2 is the best thing that happened to us (Model trainers and diffusion community in general). It trains really well, it adapts to my dataset the way SDXL could never do. I just wish its was using Flux 2 vae instead. Or even Flux 1 vae. Also, it's expensive AF to finetune. But it's worth it
@coyotte Is an inpainting version possible, or would Krea 2 edit have to come out first?
@abola2121 Maybe this works?
https://www.reddit.com/r/StableDiffusion/comments/1uhltc3/krea2_turbo_inpainting_with_lanpaint/
bruh
Having a lot of fun with your Krea model! It responds to my prompts fairly well and has a very pleasant image style. Not sure what the other people are on when they say it lacks details: either they are fans of overbaked ai-slop, or there's something wrong with their setups 🤔
Thank you so much, but actually they were correct - the fp8 was broken for 4000+ series GPUs (40's series and 50's series). It was caused by a bad conversion script that I've used. This issue is now resolved.
there really needs to be a tag for interracial and shemale content, so we can filter it out forever
I don't think civitai devs are reading this comment section tbh. You'll have better luck contacting them directly.
considering this is the wild west of porn you're willingly partaking in, you're acting like a an NPC on the Oregon Trail complaining life is too tough for your delicate sensibilities.
ironically for you - man up, bro.
PS - gays are way tougher than your self-projected image; they literally get slammed by other dudes while you out here gooning to wussy little hipster girls
Oh look, the pervy gooner only wants to see his own sexual preferences!
outside of uncalled for personal insults the very concept of "having preferences" seems unacceptable to some people that apparently struggle to imagine how others might be into different things and/or disgusted by certain things and prefer to avoid those
@leekstrider2250 What you're saying makes total sense, and I agree with you for the most part. However, this is a rather unusual place to post a website feature request, if I'm understanding your comment correctly.
@leekstrider2250 Nah it is a pretty dumb thing to complain about
"Block content from this user" is not what you asked for, but in the meantime, perhaps?
Nice fine-tune. Realistically, how much potential does Krea 2 have compared to other models, will a year of community dev drastically improve it, and do you think it can eventually surpass SDXL for NSFW capabilities?
Krea 2 has the best potential out of all the modern transformer-based models. It feels like a breath of fresh air to me, I am very hyped about it.
Yes, definitely. The only replacement for it would be a ~8b model with flux 2 vae, gemma4 text encoder, and a similar pretrain dataset that Krea team used. Or even an autoregressive model... Oh, a man can dream.
Gemma 4 TE? Isn't the Gemma 4 attention mechanism kinda not ideal if you want it to just act as a standalone Encoder?
🚨 Attention: There was a bug with the fp8 version (the main download file) caused by a faulty fp16-to-fp8 conversion script, which resulted in blurry images lacking detail. If you encountered this issue, please re-download the file.
Sorry for the inconvenience!
Thanks for checking on this and fixing it! Many creators would take the buzz and run. I commend you for listening to the users reporting issues and issuing a fix
Wish I had the Vram to try the newer architectures, but V9 is exceptional.
is it possible for finetuners to make a GGUF model down to 4gb? (i have found a 4gb GGUF of the main model). i have a 6gb GPU, and cant Use windows so i cant take advantage of windows memory share. so any model i do use needs to fit in those 6gb, one day i will get a 16gb GPU. but not only am i broke, but the market is the way it is
4gb quant would be Q2, a 2-bit quant. It would loose A LOT of quality, to the point where it might be not even worth it to run this model :(
But I could make one and upload it if you're willing to take a chance
@coyotte it doesn't seem too bad with the other Q2 4gb model i have found. but its not like i can compare it to any other models. if your fine-tune is better than the base model then i would say it wouldn't hurt to see if your fine-tune Q2 is better than the base model Q2
@SLACK69 aight, give me some time
This took more time than planned because I had to rewrite the quantization implementation - existing scripts made Q2 look way too lossy on the textures. The custom Q2 variant turned out great, so it was worth the wait. It's uploaded now, let me know how it performs!
Comparisons:
Int8_convrot: https://ibb.co/vvLYnwQC
Q2 (Existing methods): https://ibb.co/j94Xdh6Q
Q2 (My tweak): https://ibb.co/v44sh18Z
@coyotte Thank you, i will let you know
@coyotte i got an error when running any workflow with the Q2 model i downloaded here, ValueError: Unexpected architecture type in GGUF file: 'krea2'. and the model name doesn't match, "lustifyNSFWCheckpoint_v10Krea2.gguf" instead of "lustify-v10-krea-turbo-Q2_K-mixed.gguf". is it possible you uploaded the wrong model?
@coyotte i managed to get this result from the other Q2 model using your workflow https://civitai.red/images/136424860
@SLACK69 As for the different name - that's fine, civitai renames the model files when you download them.
As for the file - yes, it was the correct one, BUT it was the correct one for the implementation of GGUF loader I was using. I've reuploaded the file, please try the new one, it should work with city96's GGUF loader (apparently it uses "qwen image" architecture prefix for Krea 2, lol). I've just tested it on his loader.
@coyotte now it works like a charm, thank you so much
Could you please tell me how fast Krea 2 fp8 generates images compared to z-image-turbo fp16?
On my setup (RTX 3090ti), 832x1216 resolution, 8 steps:
Krea 2 int8_convrot: 4 seconds per gen. Krea 2 fp8: 7 seconds per gen.
Z-Image Turbo FP16: 3 second per gen.
@coyotte Thank you very much for your reply.
i did a quality comparison with same seed/lora between krea 2 and ZIT, krea beat it every time, ZIT images were far less coherent. Krea's worth the extra few seconds.
Apparently the settings I was using for the Krea 2 preview images weren't the best 💀 With a better workflow the images look stunning with almost no VAE artifacts present.
Currently working on a new WF for the Krea 2 version.
8 and 9 really benefit from the FreeU v2 node in comfy to vastly improve initial generation. V8 (1.15,1.35,0.9,1.0), and V9 likes (1.15,1.30,0.9,1.0). I will mess with B2 from time to time in a similar way you might with CFG, but those values I just mentioned are pretty solid and combined with CFG 3.5, dpmpp_2m_sde and simple/sgm_uniform for wider angle or karras for up close up detail is really the sweet spot. You do less re-rolls and really don't need any additional refinement or fixes to be honest--just run that same seed again with a prompt change like appending numbers or something to do a variation to resolve any anomalies. I don't know what I'm to be honest but I went on a mission trying to get initial generation to be more inline with less capable models that "snap better" but offer less variety and concepts without enlisting dmd2 and killing variety, and that is where I landed. Figured I would share as it really makes a huge difference in my experience, so I didn't want to keep it a secret.
Do I need the Qwen text encoder and Qwen VAE to run this, or does it work without it? I have both Forge Neo (latest) and ComfyUI (through SwarmUI). I was also wondering if you had any news on an update for DamnIllustrious (no hurry, just curious).
Yes, you need both qwen3vl_4b text encoder and Wan 2.1 VAE (you could use Qwen Vae, but it's noticeably worse than Wan's).
I am still working on Anima version of Damn. If it ends up looking real enough, I might even release it under "Lustify" name. But yeah, that thing is still in work.
@coyotte Great, thanks for the info and update.
@coyotte turned out this nuance about Wan 2.1 VAE is really important. Sometimes i got weird artifacts after upscale while was using Qwen Vae, and now just tested with Wan 2.1 VAE and no any artifacts - just obviously better. Thanks!
@msux glad i could help
@coyotte
Is there any chance that this entire workflow will fit into 16GB VRAM? this is the only thing that annoys me a bit.
The recommended WF for the Krea 2 version:
https://civitai.red/models/2503119/lustify-workflows-krea-2-sdxl?modelVersionId=3123550
The GOAT 🔥
The pictures on krea 2 gives usually a slightly yellowish tint. Also the quality sometimes feel very slightly lower than basic diff. model with nsfw lora. Theses are my firsts impression with comparaisons on same seeds for each
EDIT : ITS WAY MORE CAPABLE FOR NSFW, STILL A HUGE IMPROVEMENT
Thank you for the feedback!
As for the slightly yellowish tint - yeah, I guess it can lean that way in some lighting conditions. I didn't feel it was prominent enough to warrant a change, but if it does get a bit too warm for your taste, the model reacts very well to "cool color palette" in the prompt. Alternatively, adding the in-built color correction node in your workflow will do the trick instantly without adding any noticeable inference time.
Glad to hear you're finding it so much more capable for NSFW!
@coyotte I'll get a look at it thanks, just shared some insanes pics native 4k
@coyotte I suggest you trying euler_ancestral beta57 10 steps
Amazing model, best yet on offer for Krea 2 and well worth the buzz. That said, would you consider releasing weights in bf16 for training on top of?
Thank you for the kind words! The fp16 in the optional files is actually bf16, I just messed it up when I was renaming the files. I will reupload it with the correct naming.
@coyotte you did comparison tests on raw and turbo fp16? I did several tests when the model released, using the official models released as they are, safetensor, no quants, gguf etc...
and I find that RAW is better with anatomy than Turbo, even with speed lora RAW still was the winner in my eyes, plus I combined at least 5 Loras with strenght 1 and the model still didnt fried the image.. a small little little issue is that in some prompts/seed, and kind of rare tbh, the eyes was generated with bad quality.. but there is a simple fix.
The fix was simply generate natively in high res, 2K res and forget about the classic SDXL resolutions.
Since then I just use RAW at 2K res, very high quality on the base model.
a plus for extra quality that you could do on your workflow, is a 2 stage ksampler, something similar to wan2.2 14b. Gen the first pass without turbo lora and complete with turbo lora. And turbo lora on raw, at least for me the sweet spot is 0.90. The results are AMAZING!
I do not have points to download and test myself, that's why I'm asking... I'm very curious about it... knowing your finetunes since the very first release of sdxl were in high quality.. with krea, the thing must be the best open source model we have so far.
Man i'm at a point where i just can't decide between lustify sdxl v9, krea2 Q2 4gb and Q4 7gb. I have a 8gb vram so idk if i can run krea2 models and if they look better than sdxl. I also want to know how character training on krea2 works. I really need some help here, should i stick with the sdxl model or give it a go to krea2, would be grateful!
Personally, I have no reason to go back to SDXL ever again. The prompt adherence is on another level, and the amount of detail in a first pass alone is way higher than SDXL could ever achieve. It also adapts to my dataset so well that all those months of manual curation have finally paid off - the women look really pretty 90% of the time (and for the other 10%, they’re still pretty 😁).
And yeah, the character loras trained on base Krea 2 work just fine with my version, it's not the same SDXL situation
@coyotte Amazing. Will give the Q4 7gb one a try since i have a 8gb vram but idk if i can able to run it because these new models have very heavy text encoder models like 4gb and above.. so running them both would absolutely exceed my gpu limit.
@chrisss1 you can always offload the text encoder to the regular ram
@coyotte Oh didn't know this one, i need to search for it!
@coyotte i need to do anything special?
@chrisss1tell me what u find plz
@maybemalice What are you using? comfyui or forge neo? idk anything about comfyui so if you use that one, i can't help.
Not too impressed with Lustify-V10 unfortunately. The outputs allude that the data set consists mainly of Reddit images, that are often over-edited and lack the appropriate skin details as a result. I'm having to use 3rd party schedulers & samplers to avoid the Flux look, which is very noticeable with close up shots involving flash. It would benefit from exposure to more images that aren't edited and predominantly shot on newer mobile devices. The composition reflects this as well, especially in interior environments. Overall, the model preforms as a Top by All Time Subreddit model. I'd swap out some of the training data with candid images, older model mobile shots, & introduce raw DSLR shots in there too. Currently, its not too flexible.
Hey, thanks for the review!
It's fine if you aren't a fan of the aesthetic - after all, it's all a matter of taste. You might like future versions, as I'm constantly tweaking things.
Funnily enough, I've had people in my Discord channel having trouble making it look amateur-ish because it kept giving them professional-looking images (so, a 180 from your issue) xD. So I guess it might be related to the way you prompt? Personally, I can get any style I want, it just requires different approaches:
1) For pro high-aesthetic images, besides the subject, attire, and action, you want to describe the color palette and the vibe. Also, adding one or multiple tags from this list helps: "high aesthetic", "shot on Fujifilm GFX 50R", "shot on Fujifilm GFX 100", "helios 44", "depth of field", "bokeh", "volumetric lighting" or just describing the lighting in general. When you say it needs the introduction of "raw DSLR shots", you aren't entirely correct - those are already present.
2) For amateur-ish images, you probably don't need to do much - the model defaults to that style most of the time anyway, since it's the baseline aesthetic I chose for it for now.
3) And if you want it even more raw/amateur, using something like "candid amateur photo, iso grain" does the trick.
As for skin textures - I'm pretty convinced that we either have very different preferences, or you're using the model incorrectly, because even the default Euler + Beta 8 steps yield great results in that regard. Please test my recommended workflow: https://civitai.red/models/2503119/lustify-workflows-krea-2-sdxl?modelVersionId=3123550
If you still have problems with skin texture after that, please send me an example of what you're seeing. There was an issue where the fp8 file would output blurred/non-detailed images for some people, but it was fixed a while ago. Make sure you aren't using an old file, or try int8_convrot just in case - it outputs better images faster anyway.
@coyotte Apologies for the confusion. Yes, going for the amateur-ish aesthetic is hit or miss due to the model being inclined to output the "Over-edited" Reddit look I mentioned, especially when prompting for amateur photos. Coincidentally, I managed to generate multiple amateur photos with your model last night, so it is capable of these outputs! I appreciate the thorough response, my intentions are always to provide feedback for consideration during training for future versions. Despite having a background in fine-tuning checkpoints, I lack the resources required to do so for these larger models. If you find yourself struggling to source ethical training data for candid/amateur/vintage shots feel free to reach out directly at any point. Thank you!
Hi, does anyone know how to train a LoRA in OneTrainer or AI Toolkit using Lusitfy's RAW? Thanks!