DigiXmas - Stable Diffusion 3.5 L LoRA
Step 1 in making a LoRA is gathering my dataset. You can pull from google, Pinterest, make your own images or any other platform. As you download your images, you can rename them starting with 0.
I honestly don't have a good idea of how many images you need. I've used anywhere from 1 to 200. Depending on where you are doing your training will probably determine how big your dataset is.
Step 2 is resizing; it is recommended you resize to either 512x512 or 1024x1024. I use Bulk Resizer.
Step 3 Generate Captions. Most sites will have a caption generator built into the training app or you can use another tool to generate them. JoyCaption, FLUX Prompt Generator . These are two tools you can use to make quality captions.
Step 4 I use Image Dataset Manager to organize and edit the captions. And it will take the folder and save it as a zip file.
The next part is the hardest part of training a LoRA. Selecting your settings. On Tensor.art you can choose between Basic Mode and Advanced Mode.
With Basic Mode the only choices you have are to choose your model, Epoch and Repeat. You can also choose your Trigger Word.
I always use Advanced mode. I use ChatGPT to get my settings. I have been making Stable Diffusion 3.5 L LoRA and Stability AI posted an article with good info about FineTuning and LoRA training. I provided that and a few other documents to ChatGPT. Then I tell it what settings I need and how many images and what kind of LoRA I am making. Lastly in the prompt I tell it to use the provided documents and the internet to come up with the settings and to provide me what resources that it used.
Now for the settings.
Your setting will be determined by what you are looking to train, a Character, a Concept, or a Style. A character is pretty obvious, but it is a specific person or character, realistic or anime. A Concept is Objects, clothing, anatomy, poses, etc. Finally, a Style, a time period, art style, or general look and feel.
Your Epoch, Repeats and Image number will determine your Steps.
On my DigiXmas LoRA the settings were.
Repeat:8
Epoch:5
Save Every N Epochs: 1
Clip Skip: 1
Text Encoder learning rate:0.00001
Unet learning rate: 0.00015
LR Scheduler: cosine_with_restarts
Optimizer: CAME
Network Dim: 16
Network Alpha: 8
Gradient Accumulation Steps: 4
Keep n tokens:1
Noise offset: 0.05
Multires noise discount: 0.3
Multires noise iterations: 3
