experimental Flux dev version, 1286 images, joycaptioned, command line:

accelerate launch  --mixed_precision bf16 --num_cpu_threads_per_process 1 flux_train_network.py --pretrained_model_name_or_path /workspace/train/flux1-dev.sft --clip_l /workspace/train/clip_l.safetensors --t5xxl /workspace/train/t5xxl_fp16.safetensors --ae /workspace/train/ae.sft --cache_latents_to_disk --save_model_as safetensors --sdpa --persistent_data_loader_workers --max_data_loader_n_workers 2 --seed 42 --gradient_checkpointing --mixed_precision bf16 --save_precision bf16 --network_module networks.lora_flux --network_dim 64 --optimizer_type adamw8bit --learning_rate 0.00005 --network_train_unet_only --cache_text_encoder_outputs --cache_text_encoder_outputs_to_disk --highvram --max_train_epochs 16 --save_every_n_epochs 1 --dataset_config /workspace/train/urushihara.toml --output_dir /workspace/train --output_name urushihara3 --timestep_sampling shift --model_prediction_type raw --guidance_scale 1.0 --loss_type l2