ai-toolkit lora config

---
job: extension
config:
  # this name will be the folder and filename name
  name: "<CONCEPT_NAME>"
  process:
    - type: 'sd_trainer'
      training_folder: "output"
      performance_log_every: 10
      device: cuda:0
      network:
        type: "lora"
        linear: 16
        linear_alpha: 8
      save:
        dtype: float16 # precision to save
        save_every: 250 # save every this many steps
        max_step_saves_to_keep: 5 # how many intermittent saves to keep
        push_to_hub: false #change this to True to push your trained model to Hugging Face.
      datasets:
        - folder_path: "G:\\<PATH_TO_YOUR_FOLDER>"
          caption_ext: "txt"
          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
          shuffle_tokens: false  # shuffle caption order, split by commas
          cache_latents_to_disk: true  # leave this true unless you know what you're doing
          resolution: [ 1024 ]  # chroma enjoys multiple resolutions
      train:
        batch_size: 1
        steps: 8000  # total number of steps to train 500 - 4000 is a good range
        gradient_accumulation: 1
        train_unet: true
        train_text_encoder: false  # probably won't work with chroma
        gradient_checkpointing: true  # need the on unless you have a ton of vram
        noise_scheduler: "flowmatch" # for training only
        optimizer: "adamw8bit"
        lr: 4e-4
        content_or_style: "balanced"
        # uncomment this to skip the pre training sample
        # skip_first_sample: true
        # uncomment to completely disable sampling
        # disable_sampling: true
        # uncomment to use new vell curved weighting. Experimental but may produce better results
        # linear_timesteps: true
        ema_config:
          use_ema: true
          ema_decay: 0.99
        dtype: bf16
      model:
        name_or_path: "G:\\<PATH_TO_CHROMA_MODEL>\\chroma-unlocked-v33.safetensors"
        arch: "chroma"
        quantize: true  # run 8bit mixed precision
      sample:
        sampler: "flowmatch" # must match train.noise_scheduler
        sample_every: 100 # sample every this many steps
        width: 512
        height: 512
        prompts:
          - "<CUSTOM_PROMPT>"
          - "Example Prompt: xyhr, A middle-aged h3m7n, in a room with a folding screen, a painting, and a chair." # Here I have 2 trigger words, 1 for main concept or lora and the other which is the specific detail I try to train into a word
          - "Dog on a motorcycle" # Used to determine if original model gets overtrained
        neg: ""  # negative prompt, optional
        seed: 42
        walk_seed: true
        guidance_scale: 4
        sample_steps: 25
# you can add any additional meta info here. [name] is replaced with config name at top
meta:
  name: "[name]"
  version: '1.0'