# Starting configuration for FLUX.1 dev, not schnell or FLUX.2.
# sd-scripts 690ea7f96c23182352ec63def76d431c6120bd2f
# Run from sd-scripts with flux_train_network.py --config_file PATH.
pretrained_model_name_or_path = "/workspace/lora/models/flux1-dev.safetensors"
clip_l = "/workspace/lora/models/clip_l.safetensors"
t5xxl = "/workspace/lora/models/t5xxl_fp16.safetensors"
ae = "/workspace/lora/models/ae.safetensors"
dataset_config = "/workspace/lora/config/flux1-dev-dataset.toml"
output_dir = "/workspace/lora/flux1-dev/output-v1"
output_name = "mug-flux1-dev-v1"
save_model_as = "safetensors"
save_precision = "bf16"
network_module = "networks.lora_flux"
network_dim = 16
network_alpha = 16
network_train_unet_only = true
learning_rate = 0.0001
optimizer_type = "AdamW8bit"
lr_scheduler = "constant"
max_train_steps = 1000
gradient_accumulation_steps = 4
max_grad_norm = 1.0
mixed_precision = "bf16"
sdpa = true
gradient_checkpointing = true
fp8_base = true
blocks_to_swap = 18
guidance_scale = 1.0
timestep_sampling = "flux_shift"
model_prediction_type = "raw"
cache_latents = true
cache_latents_to_disk = true
cache_text_encoder_outputs = true
cache_text_encoder_outputs_to_disk = true
vae_batch_size = 1
max_data_loader_n_workers = 0
save_every_n_steps = 200
save_state = true
save_state_on_train_end = true
seed = 42