This is it. The one you've been waiting for. We've retrained SDXL with 1.3M professional and amateur pornographic images. This model does softcore, hardcore, and everything in-between.
The model seems to do best with resolutions 832x1216 or 1216x832. My settings:
Sampler: dpmpp_2m_sde
Scheduler: Karras
Steps: 30
CFG: 7
The model isn't perfect. You'll get some strange anatomy at times. Experimentation says this is the result of TE2, which was likely trained without much NSFW content. Still, you can get great results with a bit of cherry-picking.
Recipe:
In each step, TE weights were frozen - this model modifies only the unet. Machine captioned images were used until the last step.
Retrained SDXL 1.0 on 1.3M images for eight epochs
bs=80, lr=1e-05, optimizer=AdamW8Bit, EMA=0.95
Finetuned on 50k images for ten epochs to help do some rebalancing
bs=80, lr=2e-06, optimizer=AdamW8Bit, EMA=0.95
Aesthetic training on 30k images for 14 epochs
bs=1, lr=1e-05, optimizer=Adafactor, no EMA
Final alignment using 600 hand-selected and captioned images
bs=1, lr=1e-05, optimizer=Adafactor, EMA=0.95
Feel free to make your own merges and finetunes based on this checkpoint for non-commercial purposes. I ask that, if you do, you include a recipe with a similar level of detail as what I provided above. There's too much secrecy around this topic.
I have made checkpoints from step 2 & 3 in the recipe above available. These are not recommended for general use - they do not make good images out of the box. They may be useful for model mergers and finetuners.
Description
This is the version that corresponds to step 3 in the recipe. It does not have the final alignment and is not recommended for most users. It's available here for model mergers and trainers in case they want to do their own refining.
FAQ
Comments (52)
Trust me you don't want to prompt Chubbynator with the checkpoint without using any negatives or other words.
lol
I made a bunch of sample images in a post:
https://civitai.com/posts/2587004
I used simple prompts, and no cherry picking. As well as the recommended settings.
The color is really like an old paper photo that got left out in the sun. Not ready for a daily use for me, but it does do pubic hair really well.
agree :(
Getting really good results from v0.9 unrefined.
This is by far the best non-Pony model for this type of gen.
Also, this isn't working incredibly well with ControlNet, likely because of differences in training. Comes with the territory, I suppose.
extremly slow and kinda rough. For some reason, it takes ages to render 1 pic, with recommended settings, compared to pony xxx nsfw, also my own trainerd loras, look bad with it. Any tips, what i might be doing wrong?
se fooocus, it helps, also put your lora trigger at the start
Used 0.9 to do a quick custom finetune, just as an overnight test. 1,350 image dataset, 27k steps (yeah, too high). So far very promising results, with some of the lowest losses on the graph.
Anatomy is much more consistent than training with base SDXL. And it doesn't seem to be much more difficult to train than base SDXL.
Thanks for making this model and releasing information about how it was trained. The community needs more stuff like this.
The 6.62GB Version downloaded from the main blue download Button above doesn't work in the Draw Things App (MacOs). Pictures are like bad drawings. Is VAE included? or any other recommendations for settings when import into Draw Things App? Normally other models I have downloaded from Civitai work fine...
I was coming to check out the model page due to this same issue. Scripted this with 10 dif prompts, w/ and w/out loras and 4 samplers and not a single good image. Sounds like its something DT specific?
i can't get it work either, i've tried different import model settings but no change. It creates a great picture....and then badly paints over it? its a weird effect. I've tried all the samplers and nothing really makes it go away.
This stands out of the crowd as a unique and vivid AI generator. After seeing countless similar AI generators, this model is unique and special.
How to prompt specific models from specific sites? Is there a particular format to do that?
This model got crapped on by timing. It came out during the pony fervor. It's probably the best true photoreal model for SDXL out there. It should have beat all the other models like SDXXXL and ALL of the 1.5 models.
I hope this doesn't deter the creator from working with SD3. You are doing God's work.
All the time and expense to produce this but no prompt guidance?
Wish the person responsible would check in once in a while. Hopefully they are gonna do more training?
NGL I'm liking this model, it's better than SDXXXL for me, but the only issue is the weird "Old photo" colors it seems to produce consistently, is there any way to avoid this?
This is about the 20th model I've ever tried and after using all the recommended settings and trying out some prompts this gave me the weirdest results of them all.
Great job! Also tnx for the RAW and unrefined versions. Do you know where I could find more unrefined or RAW SDXL models?
just curious to know why you find interesting the unrefined one? is more "pure" maybe works better with merges?
@p_p Seems to work better if you would merge it with instruct pix2pix or other models if you would not want to have a big impact on the prompt style.
Could you please release the image captions?
One flaw in this model, I believe, is that it produces lots of cropped images. I'm guessing it was trained on a bunch of images which were processed in bulk and cropped to fit some size, and so the model that was trained on images cropped in the wrong places tends to produce such images.
I havn't noticed this. Do you have an example?
@erotes_anteros Any chance of a V2 of this? This is one of the best models out there.
kickass model
i'm just coming back because this model is wonderful.......i'm mindblowed after every prompt even sometimes i found myself like the result is too much for me. This model show so many strenght : realism, fidelity, versatility, prompt adherence, exceptionnal tone and colors...
One of the all time best. I tend to use it in a live mix with other models. I've been getting good results with 1328 x 1024.
Any good mixes you might post to civit plz?
How to do live mix?
Really hoping OP can get to a V2 at some point.
That would be amazing, wouldn't it
Insanely good! I spend a lot of time on creating merges (maybe I will attempt training myself some day) and this serves me very well for my absolute favorite NSFW merge: 0.5(anterosXXXL_v10) + 0.5(Iris_Lux_v1244_lim) + 0.5(pyrosNSFWSDXL_v05 - sdXL_v10VAEFix)
Do you plan to upload that merge?
@P_Universe its such a simple one, I am not sure if that warrants an upload.Also the Irisl Lus 1244 i snot a public model. You can get it on Patreon. But there is a an older version available here: https://civitai.com/models/201287/iris-lux-polyvalent-prototyperealisticsfwartnsfwpornno-refiner-needed
@kunde2 but 1051 is and uf I remember the last one he upload it for free on hugging face,but there are fi example models like bigasv2lustify that is a merge published, you should try it, it would help us a lot
@P_Universe OK, convinced, I will do it (merge with the Iris Lux that's available here though) :) But not super soon. I will let you know here though.
@kunde2 leeets gooo, much appreciated
I used the older version and wow - excellent recommendation
@kunde2 Hi, any news about the new NSFW model?
imagine merging this with bigasp
@smolasd unfortunately, bigasp's understanding and intelligence takes a hit as a result. Evidently this model was only partially trained and uses standard SDXL word understanding or something?
Wow I never realized how awesome this model is
@P_Universe Is pretty goot, but bigasp v2 is probly better.
yeah, it could use some polish, but it's trained on a large unique data set and has brought a lot to many other models.
Alot of the samples where make with a lora called 8SL. Would be nice if somebody uploaded it.
ANy good news for this model.
Expecting some progress to your great work!
FLUX maybe
Flux would be GREAT!!!!!
Flux would be GREAT!!!
Do you need the SDXL 1.0 base model for this SafeTensor to work? or is it the full model?