The base model already knows a lot out of the box. Poses, lighting, skin texture, it handles most requests without any help. But when you push it into detailed adult work it still slips on anatomy, positions, concepts and the small details that make or break an image. That is what this LoRA fixes.
Mystic XXX V1 adds a focused layer of NSFW training on top of Moss, built from a large set of images with detailed captions covering many positions and body types. The result is cleaner body structure, more believable skin and material detail, and concepts that land much closer to what you actually wrote in the prompt. It shines most when your prompt describes exactly what is happening and where, which is precisely where the base model used to wander.
This is only version one. I still call it experimental because I keep testing edge cases, but the core work is solid and usable today. Plain English descriptive prompts get the best results, with the usual caveats for anything early stage.
Description
V1 use with STR 1, all images have the simple wf included
FAQ
Comments (48)
kudos for making it a lora instead of an unneccessary checkpoint!
No problem you are a real dude!
Wonder if this fixes rimjob/anilingus, out of all the stuff that's the one that's giving me problems
@alcaitiff Noice
Greatt work!! This is what everyone was waiting for! The BASE model on its own already offers a high level of realism and allows for everything related to NSFW content, completely uncensored. But when you combine it with a 'general NSFW' LoRA like this one, it takes things to a whole new level of control and realism! Please keep updating it with even more hardcore versions, featuring a greater variety of poses and styles.
A tip for everyone: DO NOT use the turbo (SFT) model version if you want maximum realism and a broad range of outputs, use the BASE version in combination with the 8-step LoRA base version! The difference is like night and day! I’m achieving an absurd level of realism, almost indistinguishable from real life by using the BASE VERSION with these settings: (model: 'moss-image base int8' - 8-step LoRA for the 'base version' with strength between 0.65 - realism LoRAs like 'lenovo ultrareal' or 'oldnokia ultrareal' or 'Smartphone Snapshot Photo Reality ' and NSFW LoRAs to refine poses, penis and pussy anatomy, such as 'mystic xxx' / er_sde / bong_targent or sgm_uniform / CFG between 2.6 and 3.2).
And another very important tip: add tags like these to the end of your prompt to boost realism: intimate 2010s amateur smartphone snapshot, candid smartphone photo, compression artifacts, low dynamic range, candid phone photo, natural skin texture, skin pores, skin tones, fine body hair, perspiration, uneven sharpness, digital noise, JPEG compression, sharpening halos, imperfect white balance, natural daylight enters from the right side, producing warm clipped highlights across exposed skin and deeper noisy shadows toward the left.
I got worse results with the base model; most of the images turn out blurry.
@A1988 Dude, you must be doing something wrong, that can't be right. The base model gives me impressive results. Are you using the 8-step moss turbo LoRA (base version) along with the base model, and configuring everything just as I mentioned in the comment?
@xxxaigen It was the CFG; it seems that Turbo LoRA, with a CFG of 1 on the base model, is unable to generate good images.
@A1988 Yes, I explained it in the comment. The CFG must be between 2.6 and 3.2 (but if you’d like, you can test it with a slightly higher or slightly lower value). Just a reminder that I use the Turbo LoRA (base version) with a strength of 0.65
which sample and scheduler ,
Which 8-step lora for base model you mean? The Lodestones one (MOSS-Image-SFT-8-steps-dmd)?
@ZOMUOLK er_sde, res_2m, res_2s for more realism, schedulers sgm_uniform or beta.
This is awesome! We seriously need more general purpose NSFW LoRAs like this for good models. I’d absolutely love to see you make one for Qwen2.1 as well, because that model desperately needs more general NSFW LoRAs. RedqwV2 is a good start, but it’s still somewhat limited when it comes to pose variety and hardcore acts. Since SNOFS has gone silent, it’d be amazing to see other creators step up and fill that gap. Keep up the fantastic work!
Thanks! Any chance for a Qwen Image 2.1 version?
, this guy didn't charge any buzz at all for a porn lora!
This is the way.
it's killing the rest of us creators giving away for free. I guess the money is in the online generations then
@coachbate bye nobody will miss ya
he's a hero. a true gem. people should support these kinds of creators, not the one's who gatekeep ,mediocre models & loras behind a paywall.
@rtrash1440 exactly why the internet is such a terrible place. just uncalled for. gross
Am i missing something? Why is this creator crapping on another?
As it should be 💪🏻💪🏻 MysticXXX lora's are the best on the 'net.
@aphoticroots you're not missing anything. one creator who gates mediocre models behind a paywall shitting on another creator who generously shares an awesome model for free.
@coachbate cry me a river - youre asking for money for stuff other people offer for free and likely better done too. nobody wants this hobby to become another money sink, grifter
Thank you. Much appreciated.
What's the verdict, Krea 2 vs MOSS?
Moss for me it performs better and messes a lot less on minor stuff in the image, krea does bit better on the main subject, but moss does well in anatomy and such for me. My biggest cavaet is that I dont like the faces, krea2 did good faces but it was a bit tricky to get good looking faces, but blooming girl lora fixed most of it for me, just waiting on lora like that for moss now, and snof is also there.
MOSS does a way better job: Especially in lightning, details and also prompt following. Krea2 can do better on a more natural look and there are more image editing capabilities with krea2.
@jayjay881 :nod:
Moss is way more actually realistic than Krea 2, holy cow, but at the point it's also a huge headache since it's badly un-optimized out of the box and there are like 15 different things you can do to make it better. In contrast Krea 2 basically worked immediately and all depended on the quality of the model mostly. Not here. First the models are different... by a lot, Solordz seems to be giving the most actual realism while something like Moody goes far from it. I'm sure that won't be for long, the point is they really differ.
Then people are publishing checkpoints sometimes not even telling us if they included the 8 step Lora by Lodestones. If the Lora is included CFG1 should be used, otherwise CFG 3 or so.
The ER-sde sampler seems to be a lot more interesting than Euler, and Beta scheduler might be overkill for it. Model authors seem to recommend Er-sde Simple or Er-sde Beta.
So you have to figure out the steps (25 or 8, 50 is not really worth it), CFG, and more.
Then a lot of models will go all in on anime if you don't prompt otherwise. You can change that by prepending "35mm photograph..." to all your prompts but then some models become blurry a lot. You can then fix the blur... and so on. Is it better than Krea.... hell yes.
Please do for qwen imagen 2.1 🙏
Big fan of your work, thanks for making this free for creators!
qwen image 2.1? you are the best creator
Qwen 2.1 needs more loras
Anyone got a good moss image workflow with this as a lora?
Anyone has. We are - not yet 😁 Here is very basic workflow: https://civitai.red/images/146602750
Right click --> Save image as --> Load this image in ComfyUI to get the workflow.
Also, we have workflow from MysticXXX author. Look in attached files - download
@xNzX thanks!
Here is i2i editing workflow. Make bounding box and description for that. Run generation at denoise 0.8 - MOSS inpaint what you want 😁
Workflow: https://civitai.red/images/146953033
Right click --> Save image as --> Load this image in ComfyUI to get the workflow.
Good? No. I use a simple crappy WF and still get good results.
@InsidiousOne i tend to get a lot of disfiguration and jumbled bodies
@CaptainCrunch89 Try lowering the resolution using the exact same prompt etc. Start by using something under 1MP to see if the content works. And make sure each dimension is divisible by 16. Then work your way up. Sometimes I have a great prompt that just falls apart when pushing the envelop. Sometimes I can do 10 prompts at 1920x1080 and all of a sudden the next keeps failing until I drop it down to 1280x720 and then it'll work with no changes to the prompt. Good luck.
Can the base model be used?
Can you make one for Qwen 2.1?
Unchained is done for qwen 2.1
@alcaitiff goated ty. What is a better model in ur opinion? Qwen 2.1 or Moss?
@EchoSingularity you are welcome. Moss is better t2i but qwen can edit out of box each has some use.