Okay folks, this is it...SDXL with honest to gods explicit NSFW content. Big thanks goes to the original trainers of the model's recipes.
https://civarchive.com/models/140272/hassaku-xl-sfwandnsfw-alphav
https://civarchive.com/models/147480?modelVersionId=215527
https://civarchive.com/models/133005?modelVersionId=198530
https://civarchive.com/models/169868?modelVersionId=190908
SAMPLER: use any of the above's documented sampler selections. I have found good results using Restart 50 steps CFG 7-15(depending on creative details desired).
Results posted are not upscaled or re-sharped so you can get a base impression of the models results.
SDXL has a really robust text encoder. Think like a coder when prompting and you will get much better results.
You can utilize the older style of comma separating the tokens or you can do something fancy like (woman{(face[beautiful] [green eyes] [freckles]) (dress[silk] [transparent] [burgundy color]) (body[medium natural breasts] [chubby] [mature])} The idea being to not use commas but {} to enclose the entire scene, () for subjects in the scene, and [] features of the subjects. You can potentially use CamelCasing or snake_casing to string interactions of objects together.
has pointed out that for best results you need to set a different VAE. I recommend the following https://civarchive.com/models/140686/fix-fp16-errors-sdxl-lower-memory-use-sdxl-vae-fp16-fix-by-madebyollin?modelVersionId=155933
Description
first upload
FAQ
Comments (19)
I just ran an exhaustive comparison of 55 NSFW photorealistic SDXL checkpoints looking for things like prompt obedience, anatomical correctness, and various aesthetic factors, and found that House of Freyja - Ire's Mix was the pretty clear #1.
Runners-up included Better Than Words v20 and v30, Capability XL, Luna Mia v50, and Ultrium v4, in no particular order.
Thank you for the wonderful review! I have spent the last year looking at which models perform the best at various things and have done numerous mixes with 1.5SD that were okayish to my liking...but this model just shines like gold for what I ask it to do. Thank you again for the wonderful comment! Big thanks goes to the folks that trained the model's to make this mix possible.
@IreGaddr for balance I should mention that this model has a very slight issue with the "blue orbs" which appear around hairlines, ears, or skin folds (and NSFW anatomical details) that many checkpoints experience, but it's minimal. See Realities Edge v5 for an example of a checkpoint that has them very badly, if you don't know what the "blue orbs" are. But that's literally the only issue I can identify. You're right, it really shines especially in it's prompt obedience around a standard 7 CFG.
@shapeshifter83 Yeah, the "blue orbs" are hard to get rid of :( I've noticed them more around hair lines and ears more than anything (with the occasional wtf on the NSFW). The 7CFG is good. So far some of my best image results have come out of CFG7, Restart 75 steps (per the Restart paper 66-100 is where they had the best success). Definitely seen it correct some stuff with image previews on. Like out right redraw limbs to conform to prompt around 70% generation. Yeah it's slower, buuuut I'd rather spend a little extra time running a prompt once, than 5 times to get the result I want.
@IreGaddr I just tried on comfy ui switching from the bakedvae option to the latest sdxl vae completely removed the blue bubbles on the same seed even fixed the colour
This model seems great but the colour is washed out on focus and I can't find a way to switch vae
@Vinnnn yes manually setting sdxl_vae is the best way to go. without it, the blue orbs are all over every image. With it set, you have to go looking for them to find them.
Just want to clarify for users that you need to manually set sdxl_vae for best results. "Automatic" won't do it. Without the VAE set, you'll get poor contrast and blue orbs.
I saw the problem described with betterthanwords and was about to try this model. The only vae settings i see are basically automatic or none (to use one built into the checkpoint). None did fix my issue, but are there others options for vae than this? I guess using automatic and crafting, or selecting one and renaming to match the checkpoint, would be the obvious way.
I was wrong, none did not fix the problem in betterthanwords, so now I'm back to needing a way to change the vae
Love the model so far, but I've tried using it for more NSFW stuff, like the sample images (using the CIVIT web generator), and no matter what, can't get the actual penetration/intercourse aspect, even feeding the same prompt/seed as in the showcase.
Just checked the available samplers for the model on the site generator. Looks like Restart is not a choice :(. Neither is DPM++SDE 3M. I do all my generations locally in either AUTOMATIC or Comfy (depends on what I'm doing).
they don't think it be like it is with this amazing checkpoint, but it do
Ire you need to fix the problem with where images get published, you've got an unpublished identical version that is eating everyone's images they try to post here. One of the mods confirmed this (because they can see that unpublished model). I was wondering for such a long time why nothing I would post would show up here.
Just seeing this, got it taken care of.
@IreGaddr awesome dude, i'll be posting like crazy in your gallery then. thanks very much!
Please consider putting your House of Freyja-Ire's Mix model on tensor.art because at present no one can use your model on Civit.
https://tensor.art/models/685519172039384792 here ya go!
@IreGaddr Thank you very much.
Absolutely my favorite! The one small thing I'd change is that as soon as I add a prompt for glasses the whole thing becomes a cartoony mess. 🤓
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.