config:
  name: youtube_thumbnail_lora
  process:
  - datasets:
    - cache_latents_to_disk: true
      caption_dropout_rate: 0.2
      caption_ext: txt
      folder_path: /root/lorahub/youtube_thumbnail_lora/dataset
      resolution:
      - 512
      - 768
      - 1024
      shuffle_tokens: false
      token_dropout_rate: 0.01
    device: cuda:0
    model:
      is_flux: true
      name_or_path: black-forest-labs/FLUX.1-dev
      quantize: true
      text_encoder_bits: 8
    network:
      linear: 42
      linear_alpha: 42
      transformer_only: true
      type: lora
    performance_log_every: 500
    sample:
      height: 768
      neg: low quality, blurry
      prompts:
      - '[trigger], an android robot smiling and showing a phone screen which says
        ''GLIF.APP'''
      - '[trigger], a man with a surprised expression, there is the title text that
        says ''HOW TO USE GLIF!'', there are many humanoid robots in the background'
      - '[trigger], a man in a boxing match with a humanoid robot, the title says
        ''WHO WILL WIN?'''
      - '[trigger], a huge humanoid robot with red eyes walking in a destroyed city
        with fire, the title says ''THEY ARE COMING'''
      sample_every: 500
      sample_steps: 25
      sampler: flowmatch
      seed: 42
      walk_seed: true
      width: 1344
    save:
      dtype: float16
      max_step_saves_to_keep: 5
      save_every: 100
      save_format: diffusers
    train:
      batch_size: 1
      dtype: bf16
      ema_config:
        ema_decay: 0.99
        use_ema: true
      gradient_accumulation_steps: 1
      gradient_checkpointing: true
      linear_timesteps: true
      loss_type: mse
      lr: 0.0002
      noise_scheduler: flowmatch
      optimizer: adamw8bit
      reg_weight: 1.0
      steps: 1500
      target_noise_multiplier: 1.0
      train_text_encoder: false
      train_unet: true
    training_folder: /root/lorahub
    trigger_word: youtube thumbnail
    type: sd_trainer
job: extension