config:
  name: noise
  process:
  - datasets:
    - cache_latents_to_disk: true
      caption_dropout_rate: 0.2
      caption_ext: txt
      folder_path: /root/lorahub/noise/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: 1024
      neg: ''
      prompts:
      - white noise, glitch art, [trigger]
      - distortions, surreal, ghostly face [trigger]
      - distorted faces, static, [trigger]
      sample_every: 500
      sample_steps: 25
      sampler: flowmatch
      seed: 593146
      walk_seed: true
      width: 1024
    save:
      dtype: float16
      max_step_saves_to_keep: 3
      save_every: 500
      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
      steps: 1500
      target_noise_multiplier: 1
      train_text_encoder: false
      train_unet: true
    training_folder: /root/lorahub
    trigger_word: in the style of white noise, glitchy
    type: sd_trainer
job: extension
meta:
  description: is trained on a dataset filled with white noise and glitch art, designed
    to explore what visuals can emerge from the chaos. By pushing through the layers
    of distortion, it seeks to reveal hidden patterns and unexpected beauty within
    the noise.