Kuroume style LyCORIS版本v0.1-poc (ID: 34155)

Trained on the artworks of Pixiv user くろうめ, or Kuroume (link).

I really love their art style and wanted to make some similar stuff.

This is the first time I ever train a model and I don't have a lot of experience with Stable Diffusion in general so the results are a bit funky. It's somewhat working though, thus why I am posting the model here anyway.

If you have any insights on ways I could improve it please let me know. Right below are the details on how I trained it.

Following this guide, I first downloaded all of their art pieces in 250x250 (low res but otherwise downloading in bulk would've been harder.) I then upscaled them to 512x512 with R-ESRGAN 4x+ Anime6B. Lastly I labeled the images with DeepDanBooru which I had to curate to remove bad tags like "blurry" or various character tags (made a simple Python script for that).

Then for the actual training I just used the notebook they linked in the guide, making a few tweaks to the parameters: set the training model to this nice mix I found, set keep_tokens to 0 because I didn't use activation tags (to be honest I just don't really know what it is or what it does), set the steps to 5 repeats, 10 epochs (well actually 9), and 2 batch size, using 57 images resulted in 1430 steps minus a whole epoch since I'm using the 9th version, so 1290 total steps. Didn't touch the training section, and set the LoRA type to LoCon LyCORIS. And that's it.

To use it in Automatic1111's webui you first need to install this extension. Then just like any LoRA add <lora:kuroume-style:1> to your prompt (you can mess around with the 1 to set how much the model will influence the final image).

For the example images, I used NyanMix, found it worked pretty well. But honestly most models work pretty much the same.

Disclaimer

As I said previously, I'm very inexperienced. But I also did not put a lot of effort into this version, I was just trying to get it working. Think of it as a POC for my own ability at training LoRAs. I will definitely upload a new version where I'll go more thoroughly through the whole process to hopefully get better results.

(and btw if you're wondering the name of the LoRA is the artist's romanized name)

描述:

训练词语: