I am using my own scripts with kohya, but this is the gist of the main command is:

accelerate launch --num_cpu_threads_per_process=2 "train_network.py" --enable_bucket --pretrained_model_name_or_path="{model}" --train_data_dir="{train_data_dir}" --resolution={RESOLUTION},{RESOLUTION} --output_dir="{output_dir}" --logging_dir="{logging_dir}" --network_alpha="{network_alpha}" --save_model_as=safetensors --network_module=networks.lora --text_encoder_lr=2e-5 --unet_lr=0.0001 --network_dim={network_dim} --output_name="{output_name}" --lr_scheduler_num_cycles="1" --learning_rate="0.0001" --lr_scheduler="constant_with_warmup" --lr_warmup_steps="10" --train_batch_size="{train_batch_size}" --max_train_steps="{max_train_steps}" --save_every_n_epochs="1" --mixed_precision="{PRECISION}" --save_precision="{PRECISION}" --seed="{seed}" --optimizer_type="AdamW8bit" --optimizer_args weight_decay=0.1 --max_token_length=150 --clip_skip=2 --keep_tokens="1" --bucket_reso_steps=64 --shuffle_caption --xformers --bucket_no_upscale --caption_extension=".txt" --min_snr_gamma=5 --lowram --random_crop --color_aug --flip_aug

The RESOLUTION for most of them is 512 (for newest ones are 768, tho)

network_alpha = 128

network_dim = 128

train_batch_size = 2

max_train_steps, well, someone like to save a lot of versions and hence do a lot of epochs, I compute the thing in the way that there would be only one epoch.

PRECISION = 'bf16'

noise_offset = 0.0 (or if you want one, something around 0.1-0.2)

So, I make 4 LoRAs from different parts of the bigger dataset (50 images, default steps - 30) and then I mix them with each other with mult = 0.7.

After that I test the result, if it's not good, I try premixed LoRAs, sometimes I remix them with different weights or exclude some of them, etc.

That's it.

I doubt that since the dataset pics are from danbooru