dtype: bfloat16
flux_path: black-forest-labs/FLUX.1-dev
model:
  independent_condition: false
train:
  accumulate_grad_batches: 1
  condition_type: depth
  dataloader_workers: 5
  dataset:
    cache_name: data_512_2M
    condition_size:
    - 512
    - 512
    drop_image_prob: 0.1
    drop_text_prob: 0.1
    target_size:
    - 512
    - 512
    type: img
    urls:
    - https://huggingface.co/datasets/jackyhate/text-to-image-2M/resolve/main/data_512_2M/data_000045.tar
    - https://huggingface.co/datasets/jackyhate/text-to-image-2M/resolve/main/data_512_2M/data_000046.tar
  gradient_checkpointing: false
  max_steps: -1
  optimizer:
    params:
      lr: 1
      safeguard_warmup: true
      use_bias_correction: true
      weight_decay: 0.01
    type: Prodigy
  rotation_adapter_config:
    num_rotations: 8
    r: 1
    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)
  sample_interval: 100
  save_interval: 1000
  save_path: runs
  wandb:
    project: OminiControlRotation