This SDXL Anime Checkpoint is alternative for training characters LoRA
If you have a basic plan and can't train LoRA Illustrious.
The parameters that have worked for me are:
Network Module
LoRA
Use Base Model
V2 - SDXL Anime
Trigger words
-
Image Processing Parameters:
Repeat
5 (10 repeats only if the images in the dataset are few)
Epoch
10
Save Every N Epochs
1
Training Parameters:
Seed
-
Clip Skip
2
Text Encoder learning rate
0.00006
Unet learning rate
0.0003
LR Scheduler
cosine_with_restarts
lr_scheduler_num_cycles
3
num_warmup_steps
0
Optimizer
AdamW8bit
Network Dim
16
Network Alpha
8
Gradient Accumulation Steps
-
Label Parameters:
Shuffle caption
true
Keep n tokens
1
Advanced Parameters:
Noise offset
0.03
Multires noise discount
0.1
Multires noise iterations
10
conv_dim
-
conv_alpha
-
Batch Size
4
Labeling:
(in the dataset editing window)
Auto Labeling:
Image Labeling Algorithm
wd-swinv2-tagger-v3
Batch Add Labels:
Put an activation tag at the start of every text file. This is useful to make learning better and activate your Lora easier. Set keep_tokens to 1 in Parameter Settings