I was asked to upload my LoRA creation process, and as I just made my 100th LoRA, I figured now was as good a time as any. This process creates decent LoRAs of almost any 3d videogame character.
I've summarised my process into 3 steps below:
1. Collect
First we need to get pictures of the character. To do this, boot up the videogame in question, and take 20 screenshots of the character.
For example, I want to create a LoRA of Grombrindal, a character in Total War: Warhammer 3:
To start, I take 20 screenshots using the default Windows snipping tool. Ideally you'll want pictures of the character in different environments, poses etc, but don't worry too much if this is your first time making a LoRA.
After about 10 minutes of screenshotting, I have a dataset of 20 pictures of Grombrindal that is stored locally on my PC:
2: Tag:
Now that you have your dataset, you'll want to select Train a LoRA in CIVITAI:
Select Character and give your dataset a name before selecting Next:
Upload your dataset:
Once it has uploaded, select Autotag:
Enter your character's name under Prepend Tags (Be sure your character's name is unique.
If your character is called Caterpillar or Blue Sky or something else very literal, name it Blu35ky or C@terp1llar to avoid confusing Stable DIffusion)
Once named, hit Submit to begin Autotagging:
Your data set is now captioned!
If you want better results, tag environments (i.e. desert, snow) and strange poses/angles (from behind, from side, from above etc.) Again, don't stress too much if this is your first go!
3: Bake
Option a) Bake on CIVITAI (Costs 500 Buzz)
To bake on CIVITAI, just press Next, then set the following parameters:
Hit Submit, then CivitAI will pop out your baked LORA in ~55 minutes.
Option b) Bake on Google Colab (Free, requires Google account)
To bake on Google Colab, hit Download to collect your captioned dataset as a .zip file:
Unzip the captioned dataset, then open a browser tab and open Google Drive.
In My Drive (the top level folder), create a folder called "Loras":
Inside of the new Loras folder, create a folder named after your Character (i.e. Grombrindal):
Finally, create a folder inside your character folder called "dataset":
Inside the Google drive dataset folder, upload your captioned dataset (drag and drop the .jpg and .txt files)
Once that's done uploading, go to a new tab and go to the following URL:
Ensure the following parameters on that page are filled out like so (ensure project_name is the same as the name of your character folder in Google Drive!):
min_snr_gamma: 6, set to 8 if character is more cartoony though
network_dim & network_alpha: 32
train_batch_size: 5
optimizer: Prodigy:
ensure recommended values is ticked!
Once you're done, just click the play button to run the baking code:
You'll be prompted to allow for the creation of a Colab account & give the specific instance access to your drive. If you're comfortable with that, allow it and it will begin baking your LORA! It will take ~1 hour 20 minutes to complete, so go outside and touch grass or something idk.
Once it has completed, your baked LoRAs will be deposited inside your Google Drive character folder:
Download the -10 one as it will be the most cooked, and usually the best.
Voila! You have now cooked a LoRA for free, with no hardware requirements!
If you want to look at the Grombrindal model/ training data I created via this method, you can check them out here:
https://civitai.com/models/699292/grombrindal-the-white-dwarf-total-war-warhammer
Please comment if you have any questions :)