dtype: bfloat16
flux_path: black-forest-labs/FLUX.1-Fill-dev
model:
  add_cond_attn: false
  attn_mask: false
  latent_lora: false
  mask_para:
    base_ratio: 0.6
  union_cond_attn: true
  use_byt5_mask: false
model_type: flux_fill
train:
  accumulate_grad_batches: 4
  bucket_config:
    1024px:
      1:
      - 1.0
      - 2
    512px:
      1:
      - 1.0
      - 8
    768px:
      1:
      - 1.0
      - 4
  condition_type: word_fill
  dataloader_workers: 5
  dataset:
    drop_image_prob: 0.1
    drop_text_prob: 0.1
    glyph_scale: 1
    random_select: true
    type: word
  gradient_checkpointing: true
  lora_config:
    init_lora_weights: gaussian
    lora_alpha: 256
    r: 256
    target_modules: (.*x_embedder|.*(?<!single_)transformer_blocks\.[0-9]+\.norm1\.linear|.*(?<!single_)transformer_blocks\.[0-9]+\.attn\.to_k|.*(?<!single_)transformer_blocks\.[0-9]+\.attn\.to_q|.*(?<!single_)transformer_blocks\.[0-9]+\.attn\.to_v|.*(?<!single_)transformer_blocks\.[0-9]+\.attn\.to_out\.0|.*(?<!single_)transformer_blocks\.[0-9]+\.ff\.net\.2|.*single_transformer_blocks\.[0-9]+\.norm\.linear|.*single_transformer_blocks\.[0-9]+\.proj_mlp|.*single_transformer_blocks\.[0-9]+\.proj_out|.*single_transformer_blocks\.[0-9]+\.attn.to_k|.*single_transformer_blocks\.[0-9]+\.attn.to_q|.*single_transformer_blocks\.[0-9]+\.attn.to_v|.*single_transformer_blocks\.[0-9]+\.attn.to_out)
  max_steps: -1
  odm_loss:
    modelpath: epoch_100.pt
    w_loss_1: 20
    w_loss_2: 20
    w_loss_3: 20
    w_loss_4: 20
    w_loss_f: 1
  optimizer:
    params:
      lr: 1
      safeguard_warmup: true
      use_bias_correction: true
      weight_decay: 0.01
    type: Prodigy
  reuse_lora_path: reuse.safetensors
  sample_interval: 2000
  save_interval: 1000
  save_path: output_dir