General Information:

SNOFS was trained on natural language, not tags. It will work best if you use full sentences to describe what you want.

Want to support my work or help fund the training of this dataset on other models? Join the Patreon in my profile, and if you do - thank you!

Not using ComfyUI/your inference software doesn't support lokr? I've put up a merged version here: https://civarchive.com/models/2416142?modelVersionId=2716480

By the way, if anyone out there needs an expert in fine tuning text-to-image models (especially one that's also a software engineer), reach out to me. I'm currently looking for employment.

Here's a list of some of the terms that work well:

• anus

• blowjob

• deepthroat

• braless

• cowgirl position

• cum

• cunnilingus (be specific and maybe put kissing in the negative prompt)

• deepthroat

• dildo

• doggystyle position

• fingering

• hand in panties

• handjob

• hitachi magic wand

• implied blowjob

• ipcam / nightvision ipcam

• masturbating (might want to put penis in negative prompt, or specify what she's rubbing for women)

• massage

• missionary position

• naked, nude, etc.

• penis

• pregnant (and can specify trimester)

• prone position

• reverse cowgirl position

• sex

• sheer

• snapchat (and caption/text/etc)

• selfie (and mirror selfie)

• spooning position

• strap-on dildo

• tentacles

• licking testicles

• undressing

• vagina

• wet shirt

Depending on the version, the following might kind of work:

• anal sex

• anilingus

These models are under the following license:

https://huggingface.co/Ashen3/SNOFS

Flux 2 Klein 9b V1.1:

Additional training means far less body horror, even on the distilled version (but, you know, still some there). When using the distilled version of the model try playing around with more steps, adding a little cfg, etc.

Flux 2 Klein 9b V1:

Flux 2 Klein's awesome VAE means it picks up fine details incredibly well. While it still needs more training, I have some other stuff to train in the meantime so I thought it was worth it to push this out now as it can do some things incredibly well. Expect some body horror, especially if you use it with the distilled version of the model for text-to-image. I found that perhaps using more steps than 4 was helpful with the distilled version, but I also didn't try it much. Using this with the base model has far less anatomy issues. I expect them both to improve further with more training.

Right now, for text-to-image I recommend the base model. For editing, I recommend the distilled model. Note that SNOFS wasn't specifically trained on any image pairs for editing.

Training details (skip to the version 1.3 details below if you just want to know what this model can at least somewhat do right now):

I trained this as a factor 4 lokr using AI Toolkit this time. I used AI-Toolkit because when I started the training the other options had issues with their lycoris output and ComfyUI.

I think my starting learning rate was way too high at 1e-4 with an effective batch size of 4-6 or so. I quickly decreased it but it was perhaps still too high starting at 5e-5. I'm running a different training run at 1e-5 right now and it's still learning quite quickly. I might try to further train this at a very low LR and see what happens instead of starting fresh. Note: this is probably largely because of my large lokr size. I wanted to ensure I had "room" for all of the concepts but it can make things spicy.

I think the main issue people are coming into with training both this and Z-Image are what timesteps you train on. This was mostly trained on a high shift value of 3-5 as in inference Flux 2 Klein stays above the 800 timestep mark for most of the generation and maybe does 1 step out of 50 at below 200. I found I needed to test as I went and see where the generations went wrong and try to adjust on the fly.

Version 1.3:

Further training to further refine things. This might be the last version; I wasn't really making this for myself and I'm guessing the community wants me to make something for Z-Image. I'll at least try that out once the base model is out.

Note that the list is not exhaustive at all. It was trained on natural language (and that's how you should prompt!), so many concepts are in there.

Version 1.2:

Further training, expanded the dataset even more.

Also, I see a lot of people mixing this with other NSFW general loras. I'd recommend you try it by itself first.

Note: While you can use the lightning lora with this, keep in mind it won't lead to the best results. It's great for testing prompts, but it tends to mess with anatomy, smooth out texture, and lead to less variation on the same prompt.

Version 1:

This past weekend I was gone. I decided to let my 5090 chug along making a lokr for Qwen on ~5,000 hand fixed captions on sex, nudes, and other fun stuff of hand picked images with hand removed watermarks. I wasn't expecting it to get so good so quickly, so I did a few more night's worth of training. I'll do some additional training at some point here but it's already good enough to play around with.

It can do basic sex positions, blowjobs, cum, selfies, dildos, snapchat selfies with captions, etc. Female genitals are still a bit hit and miss, male genitals aren't bad. With it being a lokr and it being trained on so many images it's wildly flexible and can be used with perfect likeness of other loras.

Note that sometimes it'll do the wrong sex position even if you name it, and I'm unsure why as the captions have no errors. It will perhaps clear up a bit with more training.

I used Musubi Tuner and it was a heck of time getting it to train a lokr. I had to use another lycoris library for it (which is somewhere in the issues on the github page, IIRC), but it's possible the main one has Qwen support by now. Here are my training settings, though note that I reduced my LR over time and I also started with sigmoid timestep sampling. I was training at 640x640 and 1328x1328 buckets:

accelerate launch --num_cpu_threads_per_process 1 --mixed_precision bf16 src\musubi_tuner\qwen_image_train_network.py `

--dit Q:\AI\Models\DiffusionModels\qwen_image_bf16.safetensors `

--vae Q:\AI\Models\VAE\qwen_vae_for_training.safetensors `

--text_encoder Q:\AI\Models\CLIP\qwen_2.5_vl_7b.safetensors `

--dataset_config S:\AI\Musubi\datasetWoman.toml `

--sdpa --mixed_precision bf16 `

--gradient_accumulation_steps 4 `

--timestep_sampling qinglong_qwen `

--optimizer_type adamw8bit `

--learning_rate 3e-4 --lr_scheduler linear --lr_scheduler_min_lr_ratio=1e-5 --lr_warmup_steps 150 `

--blocks_to_swap 25 `

--gradient_checkpointing --gradient_checkpointing_cpu_offload --max_data_loader_n_workers 2 --persistent_data_loader_workers `

--network_module lycoris.kohya `

--network_args "algo=lokr" "factor=10" "bypass_mode=False" "use_fnmatch=True" "target_module=Linear" `

"target_name=unet.transformer_blocks.*.attn.to_q" `

"target_name=unet.transformer_blocks.*.attn.to_k" `

"target_name=unet.transformer_blocks.*.attn.to_v" `

"target_name=unet.transformer_blocks.*.attn.to_out.0" `

"target_name=unet.transformer_blocks.*.attn.add_q_proj" `

"target_name=unet.transformer_blocks.*.attn.add_k_proj" `

"target_name=unet.transformer_blocks.*.attn.add_v_proj" `

"target_name=unet.transformer_blocks.*.attn.to_add_out" `

"target_name=unet.transformer_blocks.*.img_mlp.net.0.proj" `

"target_name=unet.transformer_blocks.*.img_mlp.net.2" `

--network_dim 1000000000 `

--save_every_n_steps 250 --max_train_epochs 10--logging_dir=logs `

--output_dir Q:/AI/Models/Trained/Loras/Musubi/QwenWoman --output_name WomanGirls