This is a Lora for generating the classic style chastity belt on women. I trained this Lora on the LUSTIFY V4 checkpoint using around 128 hand labled images.
Recommendations:
Use 0.9 Strength
Use ~3.0cfg
The trigger is simply `chastitybelt`
You can control the color with a prefix e.g. `blue chastitybelt`
With `transparent` make the shield transparent
Additional Information:
I personally only use bf16, the checkpoint is trained in full bf16
The model is trained with xformers attention, it seems like not using xformers attention degrades anatomy
I want to wish all enjoyers of no enjoyment and denial no fun. Please post some images to this Lora, since this is my first ever Lora and I am not sure about tagging and struggles. For me it works out of the box with LUSTIFY V4.
Description
Fixed grainyness
Cum Support
Buttplug support
FAQ
Comments (10)
Can you rename 'r32_v2_385' so I don't accidently delete this in the folder next week? How long you trained this? 154k steps will take ~3 days on a 4090 at least with my default settings. CFG 3 could be an indicator for overtraining. Sadly the metadata is not available. The belt itself shows up but the images are very distorted. Downloading Lustify v4 now since on other models I get way too much artifacts.
Thank you for your comment, the model is most likely overtrained (artifacts), but less trained versions struggled severely with the layout and clarity. The model is trained for 385 Epochs on ~400 Images hence 154k Steps (with Protigy). I utilized masked training, to counteract the overtraining. I will experiment more. Currently i am not using any regularisation images and no offset noise. I want to experiment with both of them. The model is only intended to be used with Lustify v4. The model is currently far from perfect. Thank you for your comment.
Or to frame it differently. I am still improving the model and still trying out different parameters.
@SatanicaPotestas Trying out different parameters each 3 days? Could be quite time consuming^^ Have you included close up's? When inpainting my image this did not realy get better details. I do not include reg images in my LoRA. This never worked for me...
@Lucifie No, training 300 Epochs takes around 8 hours. I use close-ups and upper/lower body shots, without faces and heavy masking. I just noticed that protigy defaults to 0 weight decay. I just started using this to fight some overtraining aswell. Non the less i am really proud of this version. The overtraining seems to be in the last 1-2 decoder modules, disabling them reduces the artifacts, but increases style bleeding and bad anatomy.
@SatanicaPotestas I have 1,45 s/it on a 4090 also using prodigy. In 8h I can do only 20.000 steps. Either you are training on small resolutions, have rented some GPU or I am doing something wrong!?
@Lucifie I am multiplying by the batch size to get the step size
@SatanicaPotestas images * epoch * steps per epoch = total steps. I can do only 20.000 total steps in 8h so I am realy interested in your settings on how you can train so fast. Is this a local setup or an online service?
@Lucifie its a 4090 with batch size 16, protigy. Bf16 all the way except VAE. Cached Latents and Gradient Checkpointing. This way i have 100% util. I dont have the it/s rn, but i will give them later.
@SatanicaPotestas o_O batch size 16! You should avoid that. The details get lost as it merges 16 images before training. I only train at batch size 1. Even at 2 you can see that first details are lost. I have done tests a while ago on this and afterwards never set it higher than 1. I rather reduce images or split the LoRA than raising this number. Cuda is only at 40-50% but I can play games during training and don't have my GPU fan at max. And with batch size 16 I think it reduces the step count by factor 16. So it's not 154k.
Batch size makes training faster but you loose quality.
Rule of thumb that works for me: 100 steps per image.
@Lucifie I started with a lower batch size (also tried 1), but the model was generalising worse. But i will try it again for v3