config.yaml

config:
  name: Scribble_v2-1024px
  process:
  - datasets:
    - cache_latents_to_disk: true
      caption_dropout_rate: 0.2
      caption_ext: txt
      folder_path: /root/lorahub/Scribble_v2-1024px/dataset
      resolution:
      - 1024
      - 512
      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: 64
      linear_alpha: 64
      transformer_only: true
      type: lora
    performance_log_every: 500
    sample:
      guidance_scale: 3.5
      height: 704
      neg: ''
      prompts:
      - 'The image is divided into two halves: on the left is a simple scribble/sketch
        of what is pictured on the right, and on the right is a close-up image of
        a humanoid robot''s head and shoulders against a gray background. The robot''s
        face has a smooth, white surface with minimal features, including expressive
        gray eyes and a subtle smile. It has a sleek design with visible circuitry
        and mechanical components around the neck and shoulders. Headphones are placed
        over the ears, emphasizing its futuristic appearance. The overall look conveys
        a blend of human-like characteristics and advanced technology.'
      - 'The image is divided into two halves: on the left is a simple scribble/sketch
        of what is pictured on the right, and on the right is a close-up photograph
        of a revolver positioned on a neutral background. The firearm is primarily
        silver with a textured black grip.'
      sample_every: 500
      sample_steps: 25
      sampler: flowmatch
      seed: 1100
      walk_seed: true
      width: 1536
    save:
      dtype: float16
      max_step_saves_to_keep: 3
      save_every: 500
      save_format: diffusers
    train:
      batch_size: 2
      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: 0.0003
      noise_scheduler: flowmatch
      optimizer: adamw8bit
      reg_weight: 1.0
      steps: 3000
      target_noise_multiplier: 1.0
      train_text_encoder: false
      train_unet: true
    training_folder: /root/lorahub
    trigger_word: '-'
    type: sd_trainer
job: extension
meta:
  description: v2 with new captions + 1024px