---
job: extension
config:
  name: "flux_eldritch-v1s8000-fcompass"
  process:
    - type: 'sd_trainer'
      training_folder: "/home/kade/flux_output"
      log_dir: "/home/kade/flux_output/logs"
      device: cuda:0
      network:
        type: "lora"
        linear: 32
        linear_alpha: 16
      save:
        dtype: float16 # precision to save
        save_every: 250 # save every this many steps
        max_step_saves_to_keep: 30000
        push_to_hub: false
      datasets:
        - folder_path: "/home/kade/datasets/eldritch/1_eldritch"
          caption_ext: "txt"
          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
          shuffle_tokens: true  # shuffle caption order, split by commas
          cache_latents_to_disk: true  # leave this true unless you know what you're doing
          resolution: # flux enjoys multiple resolutions
            - 512
            - 768
            - 1024
      train:
        batch_size: 1
        steps: 8000  # total number of steps to train 500 - 4000 is a good range
        gradient_accumulation_steps: 1
        train_unet: true
        train_text_encoder: false  # probably won't work with flux
        gradient_checkpointing: true  # need the on unless you have a ton of vram
        noise_scheduler: "flowmatch" # for training only
        optimizer: "adamw"
        lr: 3e-4
        lr_scheduler: "rex_with_restarts"
        lr_scheduler_params:
          lr_end: 0.01
          first_cycle: 250
        #linear_timesteps: true
        # ema - Exponential Moving Average (https://arxiv.org/abs/2310.13854)
        # It provides stable estimates of stochastic quantities improving the performance of the model.
        # Uses momentum information, even when on a non-momentum based method.
        # NOTE: You have to comment out this whole section to disable it.
        # NOTE: we could try BELAY maybe? 2e3 seems like a nice curve.
        #ema_config:
        #  use_ema: true # ⚠️: 
        #  ema_decay: 0.99
        dtype: bf16
      model:
        name_or_path: "black-forest-labs/FLUX.1-dev"
        is_flux: true
        quantize: true
      logging:
        log_every: 1
        use_wandb: false
        verbose: false
      sample:
        sampler: "flowmatch"
        sample_every: 250
        width: 1024
        height: 1024
        prompts:
          - "image tags: anthro, male, werewolf, wolf, standing, black fur, red eyes, image captions: An anthro male werewolf standing in a dark forest, he is holding up a yellow street sign with a black werewolf silhouette and black text saying 'Werewolf breeding area!', dramatic lighting, thunder, lightning, scary horror movie aestethics, halloween, pumpkins"
        neg: "" 
        seed: 42
        walk_seed: true
        guidance_scale: 4
        sample_steps: 20
meta:
  name: "[name]"
  version: '1.0'