General NSFW Enhancer
female body (vagina, anus, breasts)
male penis
sex positions
female solo
cum
all sorts of skin colors
** v1.3b is more sharp (layers 22-31 muted) ty @SD_STORM for the tip!
Lora str 0.7 - 1.0
Workflow in all samples (er_sde, beta, cfg 3.0, 20 steps recommended)
increase CFG to get more detail
Viggle Turbo compatible:
https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/tree/main
Viggle Lora: 1.25 with 8 steps looks good for me!
Description
FAQ
Comments (28)
It smoothing everything and skin too. need to zero layers 22-31 (textural and style info). So i made a lora moddifier.
@SD_STORM interresting ty! is that why loras break with euler/cfg1? i tried it and now even euler and cfg 1 works better but the images look now little bit to much burned/crisp
im working on modifying my dataset with more penis/blowjob sideviews and some other things i may do this for the next version!
@SD_STORM That actually helps a lot, thank you for this fantastic advice! It makes the results sharper and allows for the default Euler Simple + CFG 1 settings without loss of quality, which doubles the speed and resolves issues in i2i workflows regarding too much changes compared to the reference image.
@tawsLoRa It might be worth it to re-release your LoRA with layers 22-31 set to zero. I quickly did that using custom nodes (https://github.com/lrzjason/Comfyui-LoraUtils) but since you have made the original LoRA, you should release it yourself.
@lizardon1024 and @SD_STORM i released v1.3b with the blocks 22-31 set to 0
thx again for the info :)
Last layer in lora always goes overburned and crack all. Thats why not many character loras posted. I because of big text encoder, that makes most of work. Still have questions how to train loras on qi21, in krea2 its not a problem. Can't understant difficults of edit mode and t2i with lora. Bwt viggle 0.3 lora have more layers than common (37 layers)
You can block or reduce Blocks 6-8 so it doesn't alter the face's appearance as much in Edit Mode. I reduced them to 0.2, and it works great with the first Viggle Turbo LoRAQwen-Image-2.1-viggle-turbo-4step-lora-r64.safetensors but using 5 steps, CFG 1, er_sde, and a beta, on 0.8 strength. On my RTX 3060 12GB, it generates and edits a 2MP image in 22 seconds.
BTW, THYM for this lora.
@tawsLoRa this is the best lora for nsfw right now, i have replaced flux klein with this. request to add innie pussy in this for next version
@tawsLoRa Bro, good work! Have a question: have about anal penetration stability? Have a good samples of prompts to force anal penetration?
@MetamorphosisNaturelle you can try like "they are doing reverse cowgirl sex, the penis is inserted into her anus, anal sex,"
It's a good first version, but it changes the face from the references a bit.
@A1988 i never tried it with a images as source always T2I i may will look into this to :)
works very well in image edit too
does this version changes face?
faces change?
@RANDINOVGG not in simple edits, but slightly changes in complex, multi image or multi character edit
@laylawells43353347 how does it handle cartoon and anime?
v1.3b is excellent at keeping the character and the skin texture. Best NSFW LoRA for Q2.1 so far!
1.3b is very good!
Is 1.3b the same lora with the layers muted?
yes
This works very very well, even in image edit
Unfortunately this is not working with official turbo model giving:
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.0.img_mlp.gate_up.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.0.img_mlp.gate_up.lora_B.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.1.img_mlp.gate_up.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.1.img_mlp.gate_up.lora_B.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.10.img_mlp.gate_up.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.10.img_mlp.gate_up.lora_B.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.11.img_mlp.gate_up.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.11.img_mlp.gate_up.lora_B.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.12.img_mlp.gate_up.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.transformer_blocks.12.img_mlp.gate_up.lora_B.weight
etc. etc. etc. in output. @tawsLoRa any chance you can make this lora version for official turbo model?
it works perfectly fine with the new turbo model cfg 1 and euler 8 steps
also works with the turbo lora str 1.15 and cfg 1 and euler 8 steps
maybe you need to update your comfyui or workflow?
@tawsLoRa Found the problem, I was using the wrong model downloaded from Civit. Should be using https://huggingface.co/Comfy-Org/Qwen-Image-2.1/blob/main/diffusion_models/qwen_image_2.1_turbo_int8_convrot.safetensors this one
It works good with t2i also and with image diting also I think it's better then the nsfw lora we are using right now
thanks for actually fixing penis and vaginal anatomy. the only LoRA that works well in image edit!
It’s not suitable for editing; it changes the face significantly.
Best NSFW LoRA for Qwen 2.1 so far!
Details
Files
qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors
Mirrors
qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors
QWEN21_FemaledVaginus.safetensors
vae.safetensors
NSFW_femaledVaginus-v1_3-fine.safetensors
qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors
Femaled_Vaginus_Nsfw_Enhancer__v1_3b.safetensors
qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors
qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors
qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors
qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors
qwen2_1-femaledVaginus-nsfw-v1_3-fine_L22-31.safetensors


