V1 can be used straight at 1024x1024 or other aspect ratios and gives much better results much easier in my opinion.

__________________________________________________________________________________

First attempt at training an SDXL LyCoris. Used a ton of high quality images.

Hires. fix 1.25-1,5 starting at 512 x 768 denoising strength: ~0.4

I couldn't train on 1024x due to VRam limitations.

I am very impressed with the stability in her tattoos. No doubt due to the enhanced text capabilities.

Training command:

./venv/bin/accelerate launch ./sdxl_train_network.py \
--enable_bucket \
--min_bucket_reso=256 \
--max_bucket_reso=1024 \
--pretrained_model_name_or_path=sd_xl_base_1.0.safetensors \
--train_data_dir=/img \
--resolution=1024,1024 \
--output_dir=/model \
--logging_dir=/log \
--network_dim=30 \
--network_alpha=15 \
--save_model_as=safetensors \
--network_module=lycoris.kohya \
--network_args conv_dim=30 conv_alpha=15 algo=lora \
--text_encoder_lr=0.5 \
--unet_lr=0.5 \
--output_name=liyasilver_xl \
--lr_scheduler_num_cycles=5 \
--network_dropout=0.1 \
--learning_rate=1.0 \
--lr_scheduler=constant \
--train_batch_size=2 \
--max_train_steps=11300 \
--save_every_n_epochs=1 \
--mixed_precision=bf16 \
--save_precision=bf16 \
--cache_latents \
--optimizer_type=Prodigy \
--max_data_loader_n_workers=0 \
--bucket_reso_steps=64 \
--mem_eff_attn \
--bucket_no_upscale \
--noise_offset=0.0357 \
--sample_sampler=euler_a \
--sample_prompts=/model/sample/prompt.txt \
--sample_every_n_epochs=1 \
--network_train_unet_only \
--gradient_accumulation_steps 10