Network module
kohya-ss/lora
Network type
LoRA
Network Rank/Dimension
32
Network Alpha
16
Network dropout
None
Module dropout
None
Rank dropout
None
Weight decomposition (DoRA)
False
Rank-stabilized
False
Gradient Checkpointing
True
Network args
None
LR Scheduler
cosine_with_restarts
Warmup steps
0
UNet learning rate
0.0005
Text encoder learning rate
5e-05
Optimizer
bitsandbytes.optim.adamw.AdamW8bit(weight_decay=0.1)
Seed
1228666489
Max Grad Norm
1.0
Scale Weight Norms
None
CLIP Skip
1
Precision
bf16
Mixed precision
bf16
Full fp16
False
fp8 base
Get block weights
Epoch
113
Steps
1356
Max Train Steps
1680
Num train images
30
Num batches per epoch
12
Batch size
4
Gradient Accumulation Steps
1
IP noise gamma
None
Noise offset
0.1
Adaptive noise scale
None
MultiRes Noise Iterations
6
MultiRes Noise Discount
0.3
Loss type
Debiased Estimation
False
Min SNR Gamma
5.0
Zero Terminal SNR
False
Masked Loss
Max token length
225
Caption dropout rate
0.0
Caption dropout every n epochs
0
Caption tag dropout rate
0.0
Dataset
Buckets
True
Min bucket resolution
256
Max bucket resolution
2048
Resolution
1024x1024
Bucket 0
576x576: 2 images
Bucket 1
768x1344: 1 image
Bucket 2
832x1088: 19 images
Bucket 3
832x1152: 2 images
Bucket 4
896x1024: 1 image
Bucket 5
896x1152: 1 image
Bucket 6
960x1088: 1 image
Bucket 7
1024x1024: 3 images