2026/3/29 Updated
Introduction
This article serves as a personal memorandum of the settings and parameters that I found to be optimal for training Style LoRA on the Illustrious model.
I have researched this topic and tested these settings through actual training sessions.
I will not explain why these settings are good; if you are curious about the reasoning, please research it on your own.
(I am not a native English speaker, so please forgive any unnatural expressions.)
About Training Materials (Preparation Phase)
In machine learning, there is a saying: "Garbage in, Garbage out." As the phrase suggests, the quality of the output depends heavily on the quality of the input.
Please be particular about the quality of the images you use for your dataset.
Specifically:
・Do not use black and white images (unless you want the LoRA to output black and white images).
・Avoid low-resolution images as much as possible (since Illustrious is premised on 1024x1024 training).
I recommend strictly adhering to these two points.
Regarding the number of images, about 50 images should be sufficient for a Style LoRA. (If you don't have enough, manually duplicating images to pad the dataset is not a bad idea).
Tagging & Trigger Words
regarding tagging your training materials on Civitai: I change the "Max Tags" from the default 10 to 30 before running "Auto Label".
Also, regarding the Trigger Word: Make sure to set one!
About Steps
Target a total of around 3000 steps (calculated based on a Batch Size of 1).
To achieve this, fix "Num Repeats" to 1 and adjust the number of Epochs to reach the target step count.
Other Various Parameters
First, select "illustrious" for the "Base Model for Training".
The "Training Parameters" are as follows (only changes are listed):
・Epochs: Adjust according to your dataset size (refer to "About Steps" above).
・Num Repeats: 1
・Train Batch Size: 4 (As a result, the "Steps" displayed below should end up around 750).
・Shuffle Tags: Checked
・Keep Tokens: 1 (This should be automatic if "Shuffle Tags" is checked).
・Unet LR: 1 → 0.5
・Text Encoder LR: 0 → 0.5
・LR Scheduler: cosine (This will automatically be selected if the Optimizer is set to Prodigy).
・Network Alpha: 32 (Set so that Network Dim = Network Alpha = 32).
・Noise Offset: 0.04
Set to 0.05 for art styles with strong features (so-called "thick" or "intense" styles).
Set to 0.03 for art styles with weak features ("thin" or "light" styles).
・Optimizer: Prodigy
Now, just wait for the LoRA to complete.
Notes
When generating images using the created LoRA, try to place the Trigger Word at the beginning of your quality tag group.
(Example: (Trigger Word), masterpiece, best quality, amazing quality)
If you found this helpful, I would appreciate a Like, Bookmark, or Buzz (Tip)!