Update: fixed over-bright 8-step images
The text prompts I used during distillation never described lighting, so the 8-step LoRA learned to brighten every image toward a bright studio look, ignoring the lighting in your prompt. I re-distilled it with lighting descriptions added to those training prompts, and dark and moody scenes now stay dark. Examples: Image1 and Image2
Please re-download moss_pussy_lora_v3.0_8steps.safetensors and moss_pussy_lora_v3.0_8steps_dmd2_only.safetensors. The other two files are unchanged.
MOSS Better Nude LoRA v3.0
I realized one merged file isn't very user-friendly, so v3.0 is uploaded as three separate LoRAs that you stack yourself, plus an all-in-one file for convenience:
moss_pussy_lora_v3.0.safetensors: The pussy / body LoRA. Every face was masked out during training (face loss = 0), so it learns anatomy only and leaves faces to the base model. (阴部 / 身体 LoRA。训练时所有人脸都被遮掉了（脸部 loss 为 0），所以它只学解剖结构，脸交给底模来生成)
moss_pussy_lora_v3.0_8steps_dmd2_only.safetensors: The 8-step speed-up only. Fine-tuned from lodestones' DMD2 specifically for v3.0, so it works best paired with v3.0. (只负责 8 步加速。在 lodestones 的 DMD2 基础上专门为 v3.0 微调过，所以最好搭配 v3.0 使用)
moss_pussy_lora_v3.0_face.safetensors: Optional face LoRA. Swap it for any face LoRA you like. (only Asian faces in this lora) (可选的人脸 LoRA，可以换成任何你喜欢的人脸 LoRA)
moss_pussy_lora_v3.0_8steps.safetensors: All-in-one 8-step LoRA: 8steps_dmd2_only 1.0 + v3.0 1.0 + face 0.8 in a single file. Use it alone at strength 1.0. (8 步合一版。等于 8steps_dmd2_only 1.0 + v3.0 1.0 + face 0.8 合并成一个文件，单独使用，强度 1.0。)
Why a separate face LoRA? Because faces were masked out when training the pussy LoRA, faces now come straight from MOSS Base. That brings back face diversity: you no longer get the same face every time. But it also means there's no aesthetic push, so faces aren't always pretty. (Earlier I swapped faces with Z-Image, which pulls everything toward the same "beautiful" look.) The face LoRA nudges faces toward a more attractive look while keeping them varied. It's added after distillation, so it doesn't collapse into one identity. Use it, lower it, or replace it with your own.
Usage
Model: MOSS-Image Base
8 steps, separate files: 8steps_dmd2_only 1.0 → v3.0 1.0 → face 0.8, workflow
8 steps, all-in-one: v3.0_8steps 1.0 only. Don't add the other three on top, workflow
Both 8-step setups: 8 steps, CFG 1, euler + simple
50 steps: v3.0 (+ face if you want it) without the 8-step LoRA, CFG ~3, workflow
Credit: 8-step base distill by lodestones (DMD2).
v3.0 — 8 steps
What changed from v2.0 (8 steps) In v2.0 the face LoRA was part of the teacher during distillation. DMD distillation is mode-seeking, so it collapsed the face LoRA into one dominant face, and most generations ended up with the same face.
v3.0 fixes this:
Retrained the pussy LoRA with every face masked out of the loss. It learns anatomy and body detail only, not faces.
New 8-step distillation: fine-tuned from lodestones' DMD2 LoRA, using Base + the new pussy LoRA as the teacher (no face LoRA in the teacher).
Face LoRA added after distillation, so it shapes the look of the face without being collapsed into a single identity.
Everything is merged into one 8-step LoRA.
Result: faces are diverse again across seeds, while the genital detail at 8 steps is closer to the 50-step model than plain DMD2.
Usage
Model: MOSS-Image Base
LoRA strength 1.0. Don't add another 8-step / DMD LoRA, because it's already included.
8 steps, CFG 1, euler + simple
Credit: 8-step base distill by lodestones (DMD2)
How V2.0 8-steps version was made：
This is an 8-step version of my moss_pussy_lora_v2.0, re-distilled so the fast sampler matches the full 50-step look.
Teacher: MOSS-Image-Base with my v2.0 face + pussy LoRAs merged at 1.0, sampled at 50 steps, CFG 3, with skip-layer guidance on block 22.
Student: initialized from lodestones' 8-step DMD2 LoRA, then trained with DMD2 distribution matching for 300 student updates on ~640 images from my dataset. Built on NVIDIA's open-source
modelopt.torch.fastgenDMD implementation.Use: MOSS-Image-Base, 8 steps, CFG 1, euler / simple scheduler (the schedule it was distilled on).
Credits: Huge thanks to lodestones for the original 8-step DMD/DMD2 LoRAs (MOSS-Image-Base-8-steps-dmd, Kromoss), which this is fine-tuned from, to NVIDIA for the open DMD2 code in Model-Optimizer.
V2.0 added more sex scenes
V2.0 face + pussy merged moss_face_regface_v2.safetensors to V2.0
workflow.json attached
Check this post (https://civarchive.com/posts/31908504) for sex scenes prompts.
This LoRA was trained on MOSS-Image-Base and can also be used with MOSS-Image-SFT, but I find that the base model produces better results.
Lora strength: 1.0
Steps: 25 or 50
CFG: 3 or 4
Sampler: euler or er_sde
Scheduler: Simple or Beta
Resolution: 1024x1536 or 1536x1024
Only Asian female in the dataset！
pussy lora for MOSS image
Description
v2.0 8-step distilled (based on lodestones' 8-step DMD2 LoRA)
FAQ
Comments (7)
So 8-step LoRA act as DMD2 LoRA? And i may choose which one to use?
BTW. In your workflows LayerSkipGuidianceDiTSimple you use single layer skip at 26. Is it correct? For me - it is working perfect for almost every MOSS checkpoint. Much better than 22. But i'm still curios - what is it all about?
I fine-tuned lodestones' DMD2 LoRA against MOSS-Image-Base with my v2.0 pussy LoRA merged in at 1.0. Then I merged the fine-tuned DMD2 LoRA with the v2.0 pussy LoRA, so the workflow only needs this one LoRA (moss_pussy_lora_v2.0_8steps.safetensors). Delete the LayerSkipGuidanceDiTSimple node and set CFG to 1.0, steps to 8, and sampler/scheduler to euler + simple. Sample workflow is here: https://civitai.red/api/download/models/3434416?fileId=3331060
https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md
From the official user guide:
SLG (Skip Layer Guidance): We have implemented two variants:
Standard SLG: scale 3, 0.01-1.0.
Two-Pass Sampling SLG: We recommend uncond skip, with 22 as the recommended skip layer (0-indexed). Layers 22, 24, and 26 can all help improve image quality.
26 works best for me too.
@salttaro i understand - it is a Lodestone DMD2 LoRA with your improvements. Still, if i follow this guide almost every time i get oversaturated images with micro-deformations. Mostly with LoRAs. 50 steps and 3 CFG with Skip and Flow - lead to undesired result, most of the times. While Skip improve image a Flow distort image parts. Like on one side of the pole a face behind it have double chin and on other side second chin is disappear complettely. From time to time, especially when LoRA is used, Sampler overdraw patterns of prompted features. Like separated elements, that stand next to each other, start mix in something another, not prompted feature. Here my little investigation. https://civitai.red/posts/31967533 With fine-tuned LoRA, like yours, difference is barely noticeable. But most of the LoRAs is not even WELL-tuned. New times, new beginnings...