metadata
license: mit
This is NOT A GUIDE for LoRA training, and please NEVER treat it as.
This serves as a onging project that I share some of the personal opinions or interesting findings where I learned or discovered during my journey in LoRA training.
To start with, LoRA training could be easy and quick with simple dataset preparation and training setup; but surely that's not what I'm satisfied with.
Based on my personal experience, preparing a LoRA's dataset properly is more important than using complicated and fancy training techniques. Just like cooking, you can't make any tasty dishes using materials without good quality.
I'll disscuss several steps that I usually do when I prepare the dataset for a LoRA in order, and I'll skip the "How to Get Image Data" part since there're too many tools/methods available for you to use.
- After getting all image data needed for a LoRA, remove all none-static image files (.mp4, .wav, .gif, etc). Then change all none-.png format images into .png.
- Spend some time going through the images to see if there're any duplicated images and remove them. This will happen a lot if the dataset is large.
- Upscale all low-resolution images to at least 2k resolution. But why? Let's see an example:
We have two identical png images, one with 907x823 pixels, the other with 3624x3288 pixels using simple 4x-upscale from SD WebUI.
Now we will crop them into roughly 1k resolution using SD WebUI's _Auto-size Crop_: