Table 2 LoRA training parameter configuration

Dataset

494 high-quality images of Cantonese embroidery works

Semantic Types

kapok flowers, plum blossoms, chrysanthemums, orchids, lotus flowers, peach blossoms, roses, peonies, leaves, chicks, roosters, hens, peacocks, pheasants, mandarin ducks, magpies, sparrows, rhododendrons, goldfish, branches, lychees, peaches, pomegranates, grapes, humans, horses, pandas, tigers, dogs, sheep, squirrels, monkeys, butterflies, bamboo, banana plants, willows, maple trees, pine trees, pumpkins, palaces, gardens, temples, mountains, water, beaches

Model_train_type

sd-lora

Pre-trained Model

v1-5-pruned-emaonly.safetensors

Enable_bucket

Min _bucket_reso:512

Max _bucket_reso:1024

Bucket_reso_steps:64

Resolution

512,512

Model output

gx_lora3.safetensors

Save_precision

fp16

Rounds

12

Max Train Epochs

10

Train _batch_size

2

Learning_rate

Global:1e-4

U-Net lr:1e-4

Text Encoder:1e-5

Scheduler:cosine_with_restarts

Optimizer_type

Adamw8bit

Network

Module:networks.lora

Network_dim:64

Network_alpha:64

Network_dropout:0