Key changes:

• Removed background of each image, then:

• Removed low quality images

• - or images with small or low quality characters or style

• - or images with incorrect and bad background removal or foreground/character separation.

• epoch raised to 20

• Conv dim/alpha 8/1 for better flexibility with multiple LoRAs generation.

• Image Tags correctly generated after background removed.

Steps: 4400 · Loss: 0.393

Parameter Settings

Network Module

LoRA

Use Base Model

Z-Image - Base

Trigger words

-

Image Processing Parameters

Repeat

20

Epoch

1

Save Every N Epochs

1

Training Parameters

Seed

-

Clip Skip

2

Text Encoder learning rate

1

Unet/DiT learning rate

1

LR Scheduler

constant_with_warmup

Optimizer

Prodigy

Network Dim

128

Network Alpha

64

Gradient Accumulation Steps

-

Label Parameters

Shuffle caption

-

Keep n tokens

-

Advanced Parameters

Noise offset

0.03

Multires noise discount

0.3

Multires noise iterations

10

conv_dim

8

conv_alpha

1

Batch Size

-