Hello everyone, I have a few questions.
Is there any good Google Colab specifically for clothes training? Or If any model is the best for (stockings, lingerie, garter belts, high heels)
When making captioning, should I only leave (stockings, lingerie, garter belts, high heels, black lingerie) tags, so any other like (standing, big breast, window, mirror) not necessary?
For stockings Lora, images of full body standing woman is preferable, or closeup like a piece of garter belt or stocking is normal?
Do I need to resize images or I just can set Enable bucket?
If there are any specific settings on Koyha_ss for training with clothes?
1 Answer
Hi,
There is no Colabs for training specific concepts, i believe. You may use any decently configurable colab you like.
You should make captions for everything on your picture. If you'll do opposite, training algorithm would thing, that untagged items are parts of your training concept. It's better to use any autotagger you have for captions and then manually edit unwanted and wanted tags / captions.
Close-ups are okay, but not go to deep into this, or "shot" can be trained as well and LoRA would produce shots of close-ups too frequently.
Of course, it's better to crop your pictures to desired resolution, but you may leave it to bucketing.
Clothes are pretty simple concept. I'm using something like this:
lr / unet - 5e-4
text encoder - 2.5e-4
300 images with 1 repeat / 150 images with 2 repeats etc, you should have ~300 images per epoch (excluding batch size), 8 epochs
If LoRA changing style, it's better to lower lr / unet / te up to 3e-4 for lr / unet and 1.5e-4 for text encoder
pick batch size depending on your traning resolution and graphics card VRAM
So, you may try this.