configs/KREA2.knf.yaml

# ==============================================================
# == Krea 2 Raw — Rook LoRA
# ==============================================================

# -- Map exported subset names to source subfolders in MANIFEST.knf.xlsx.
mappings:
  felkin_anchor:
    - "0-ANCHOR"
  genitals:
    - "6-DICK"
    - "6-SHEATH"
  fullbody:
    - "0-FULLBODY"
    - "99-VERSION-LEANER"
  body_details:
    - "1-TORSO"
    - "3-HANDS"
    - "4-LEGS"
    - "5-TAIL"
  fur:
    - "7-FUR"
  head:
    - "2-HEAD"

export_sfw_subset: false

# ==============================================================
# == Images and captions
# ==============================================================

augmentations:
  mirrored_extra: true
  mirrored_transform: flip_only
  seed: 12345

image_resize:
  min_pixel_area: 512
  max_pixel_area: 768

caption:
  subject_sex: male
  add_domains_to_tags: false

# > Krea 2 uses a Qwen-style text stack, so retain the established prose captions.
caption_outputs:
  mode: hybrid_txt
  formats: [nlg]
  tag_scope: supplemental

publishing:
  huggingface:
    pretty_name: Ladybug Felkin
    version: v5.5-krea2
    optimized_for_model: Krea 2 Raw
    tags: [krea2, krea-2-raw, lora, diffusion-training, image-captioning]
    adult_content: true

# ==============================================================
# == LoRA training
# ==============================================================

training:
  enabled: true
  backend: simpletuner

  simpletuner:
    # > Keep ANIMA's intentional balance of identity, body, and detail images.
    subsets:
      felkin_anchor:
        probability: 1
        crop: false
      fullbody:
        probability: 4
        crop: false
        repeats: 1  # !! doubled once. This keeps buckets at least 4
      body_details:
        probability: 4
        crop: false
      fur:
        probability: 1
        crop: false
      head:
        probability: 2
        crop: false
      genitals:
        probability: 3
        crop: false

    # > These limits, rather than a top-level resolution alone, control Krea 2 input sizing.
    dataset:
      type: local
      dataset_type: image
      resolution: 768
      minimum_image_size: 512
      maximum_image_size: 768
      target_downsample_size: 768
      resolution_type: pixel_area
      caption_strategy: textfile
      metadata_backend: discovery
      crop: false
      crop_style: center
      crop_aspect: preserve

    # > Keep exactly two prompts from ANIMA: face identity and explicit anatomy.
    validation_prompts:
      enabled: true
      custom:
        portrait: "Rook_Kaefer, a male felkin character, close-up portrait, detailed expressive face, black and red fur markings, soft studio lighting, clean illustration."
        sheath: "Rook_Kaefer, a nude male felkin character, fully sheathed genitalia, clean lighting, neutral reference pose, detailed illustration."
        penis: "Rook_Kaefer, a nude male felkin character, erect penis is knotted and flared, clean lighting, neutral reference pose, detailed illustration."
      from_manifest:
        enabled: true
        seed: 42
        profile: nlg
      # from_prompts:
      #   caption_styles: [nlg]
      
      
    model:
      model_type: lora
      model_family: krea2
      model_flavour: raw
      pretrained_model_name_or_path: krea/Krea-2-Raw

      lora_rank: 64
      lora_alpha: 64
      lora_type: standard

      # > Save standard PEFT/Diffusers keys.  ComfyUI maps these native Krea 2
      # > Q/K/V/O adapters directly; fused QKV adapter keys would not load there.
      lora_format: diffusers
      fuse_qkv_projections: false
      flux_lora_target: null

      # > The official Diffusers base downloads on first training start.  Do not
      # > point this run at the separate local ComfyUI FP8 checkpoint.
      pretrained_transformer_model_name_or_path: null
      pretrained_transformer_subfolder: null
      pretrained_unet_model_name_or_path: null
      pretrained_vae_model_name_or_path: null
      pretrained_t5_model_name_or_path: null
      pretrained_text_encoder_model_name_or_path: null

      # > Int8 base weights with BF16 LoRA weights lower the 1024px VRAM requirement.
      base_model_precision: int8-torchao
      base_model_default_dtype: bf16
      mixed_precision: bf16
      quantize_via: cpu

      # > Omit Torch compile: it adds substantial VRAM and is not needed for batch 1.
      dynamo_backend: null
      dynamo_mode: null
      dynamo_use_regional_compilation: false

    # > Learn full-body identity and anatomy first, then expose the complete mix.
    curriculum:
      enabled: true
      phases:
        - name: focus_dick
          start_step: 0
          subsets: [genitals, fullbody]
        - name: focus_head
          start_step: 400
          subsets: [head, fullbody]
        - name: full_mix
          start_step: 600
          subsets: all

    trainer:
      tracker_project_name: kneifftools-lora-training
      tracker_run_name: ladybird-krea2
      hub_model_id: ladybird-krea2-lora

      optimizer: optimi-lion
      learning_rate: 0.0001
      lr_scheduler: constant_with_warmup
      lr_warmup_steps: 100
      seed: 42
      gradient_checkpointing: true
      max_grad_norm: 0.01
      grad_clip_method: norm

      train_batch_size: 1
      gradient_accumulation_steps: 4
      num_train_epochs: 0
      # > A 300-image set reaches roughly four nominal passes at this ceiling;
      # > choose the best 400/600/800/1000/1200-step validation checkpoint.
      max_train_steps: 1200
      ignore_final_epochs: true

      caption_dropout_probability: 0.0
      data_backend_sampling: auto-weighting

      checkpoints_total_limit: 3
      checkpoint_step_interval: 200
      validation_step_interval: 200
      disable_benchmark: false
      validation_prompt_library: false
      validation_disable_unconditional: true
      validation_resolution: 768x768
      validation_num_inference_steps: 28
      validation_guidance: 4.5
      validation_guidance_real: null
      validation_guidance_rescale: 0.0
      validation_seed: 42
      num_eval_images: 1
      validation_negative_prompt: "worst quality, low quality, blurry, cropped, malformed anatomy"

      report_to: tensorboard
      logging_dir: logs
      aspect_bucket_alignment: 16
      tokenizer_max_length: 512
      t5_padding: null

    # > Artifact-only smoke profile: prepare it, but do not launch this run.
    trainer_testrun:
      disable_benchmark: true
      tracker_run_name: ladybird-krea2-testrun
      max_train_steps: 12
      checkpoint_step_interval: 12
      validation_step_interval: 12
      validation_num_inference_steps: 12