HOW TO TRAIN LORA IN COLAB
Links:
Lora Trainer
Dataset Maker (ITS OPTIONAL IF YOU KNOW HOW TO DESCRIBE IMAGE)
UPLOADING PICTURES TO COLAB
So create "Loras" folder in google drive then create another folder should contain the character name without spaces after you created it YES create another folder called "dataset" and here you can upload images of the character.
Dataset Maker
Open Dataset Maker form the link run first cell "project_name" is the name of the first folder in Loras folder , then run fourth cell you can add to "blacklist_tags" blur, blurry background . Now run fifth cell "global_activation_tag" this is a trigger word for the lora to work you can put here for example name of the character.
Sample dataset
Lora Trainer
"Project_name"
is the name of the first folder in Loras folder
**"Training_model" **
For anime character choose AnimeFull or AnyLora
For realistic character choose Stable diffusion 1.5 or AnyLora
"Resolution"
512 is standart resolution in SD but you can choose higher to add more details for character but it will increase training time
"Flip_aug"
It will just double your images turn it on if you have less than 20 images or turn it off if you care about asymmetrical elements in your lora
Steps/How much images for good lora
So in my case around 50/80 images work the best . For example you cant use 1k images you will probably get an error "Error: Your total steps are to high. You probably made a mistake. Aborting...."
So for steps good starting point is around 2k if you have 10 images try to increase num_repeats or epochs
How to train style of artist
Insted of Lora choose LoCon
How to use Lora in colab version of the Stable Diffusion go to the output folder right click at highest epoch click Share then share again
Then make link public
Copy link and paste it in "2. Download Checkpoints, VAEs, LoRAs each url separated by a comma: Please provide links for the Civitai API, Google Drive, or Hugging Face."
Your Lora will be here if you did everything right.
If your Lora looks wrong or something like that use less weight from 0.75 to 0.85.
Or go to higher resolution choose upscaler and add your prompts to High res fix if it still looks bad your Lora is probably overcooked/overtrained
so just take a lower epoch
(MAX RESOLUTION IS 1024x1024 (not including the upscaler))
IF YOU DONT KNOW HOW TO USE STABLE DIFFUSION ON COLAB
SD COLAB GUIDE
##By Sztef🥺