Trigger words: DeltaBreastTie, bound, rope
Use Case:
DeltaBreastTie, bound, rope,wariza
DeltaBreastTie, bound, rope, from side
DeltaBreastTie, bound, rope, nipple tweak
DeltaBreastTie, bound, rope, on stomach, sex, virginal
DeltaBreastTie, bound, rope, on back
DeltaBreastTie, bound, rope, small breasts
DeltaBreastTie, bound, rope, medium breasts
DeltaBreastTie, bound, rope, large breasts
Description
FAQ
Comments (5)
Im looking toward learning how to train loras myself, what gpu do you have? I have a 5070ti and im wondering if it'll be good enough. Also do you have any tips? Especially when it comes to tagging training images
I'm using 4070ti, so there should be no problem with your hardware. But personaly I recommend training your lora on Civitai since it is faster and their default parameters are good enough for most common concept loras. A single lora takes me around 12 hrs (1000 iters) to train on my PC, but only takes around 1hr on Civitai server. The key to good loras is proper tags. Make sure your tags precisely describes the contents in the image, and does not contain things which are part of your lora concept ( e.g. I won't allow "rope" to appear in my shibari lora training set)
Also size of dataset doesn't influence your result that much according to my observation. Some of my loras are trained with merely 20-30 images, and still achieves good results. The only problem with these loras is that their transferability (e.g. trained with sitted images and ask for standing outcome) is a bit limited when trained with images of little variations.
@Apoleon I see, thank you. Is time difference the only real difference between training on civitai or on my gpu? Also i read you need to pay to train nsfw stuff, so I'd like to ask if you've paid to train any of your loras, or is it kind of a false alarm? If im very thorough tagging out stuff i don't want will it improve the quality or is there a chance itll make it worse? Also to create variations of lets say position in a certain shibari pose, for example one with sitting and one with standing, i should still tag both of these, but then include them while using the lora, correct?
@Apoleon My lora doesn't seem to be having any effect at all, I have no idea why. You said it took 12 hours on your pc, did you use kohya? If so what settings did you use?
I'm using the trigger word i trained it for, i trained it with 70 images, 0.00005 unet learning rate, 0.000005 text encoder learning rate, for 7000 steps, 10 repetitions each image, made it save each 500 steps, and neither model has any effect at all. I trained it for model stabilityai/stable-diffusion-xl-base-1.0, i'm trying to use it on waiillustrioussdxl_v150
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.