This model removes belly buttons. No keyword required.
How it was trained:
Lora overtrained on an image
Second lora overtrained on an edit of that image (in this case the belly button was inpainted out)
Repeated the above a few times
Merged all the loras trained on edited images into checkpoint_a
Merged all the loras trained on unedited images into checkpoint_b
Extracted lora from checkpoint_a - checkpoint_b
Zeroed out non-required blocks in the lora using trial and error to reduce lora side effects on generated image
Used the following guide:
https://civarchive.com/models/98334/difflorafaceshadowtweaker
https://1lemon1lemon.netlify.app/2023-06-27_difflora/
Used this for training:
https://github.com/bmaltais/kohya_ss
Used the following command for overtraining loras (paths replaced with ...):
accelerate launch --num_cpu_threads_per_process=2 "train_network.py" --enable_bucket --pretrained_model_name_or_path="..." --train_data_dir="..." --resolution="768,768" --output_dir="..." --logging_dir="..." --network_alpha="16" --save_model_as=safetensors --network_module=networks.lora --unet_lr=0.001 --network_train_unet_only --network_dim=32 --output_name="01a" --lr_scheduler_num_cycles="1" --learning_rate="0.001" --lr_scheduler="constant" --train_batch_size="1" --max_train_steps="400" --save_every_n_epochs="1" --mixed_precision="fp16" --save_precision="fp16" --cache_latents --optimizer_type="AdamW8bit" --max_data_loader_n_workers="0" --bucket_reso_steps=64 --xformers --bucket_no_upscaleUsed this A1111 extension for merging:
Description
FAQ
Comments (15)
But why ?
Why not?
Isn't it obvious? Because!
Waiting for nipple remover
I blame you for this
@m99 cool! Maybe I will get rid with nipples visible through clothes :D thanks!
Is there a keyword incase the strength is not working?
There's no keyword. The text encoder wasn't trained and the unet was made by subracting models from each other (1girl was the keyword used in the initial training but it won't likely have any affect).
Just increase the lora strength if the effect is not working. If it still doesn't work at strength 2 then the model you're using might be incompatible with this lora.
This is actually really useful, one thing AI generation commonly gets wrong is adding belly buttons to skin-tight clothing, this LORA fixes that issue.
The original motivation for this was that I'd sometimes get double belly buttons or black hole belly buttons and putting "belly button" or similar words in the negative prompt had no effect.
Thanks so much for giving a breakdown of how you did it. In your instructions where you say you repeat the overtraining of image, is each time a different set of images, or is it the same set you initially created?
Different images, although some improvement may still be had for merging loras trained on the same dataset but with different seeds.
Any chance you could make a noseless one next? Would love if you could
Need a version for SDXL/Pony models
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.