configs/F2K_9B.knf.yaml

# ==============================================================
# == Directories
# ==============================================================


# > Dataset sync derives SOURCE/, MANIFEST.knf.xlsx, and HF/<config-id>/ from this filename.

# -- Map exported subset names to source subfolders in MANIFEST.knf.xlsx.
mappings:
  # > Anchor identity examples get sampled at a lower probability below.
  felkin_anchor:
    - "0-ANCHOR" # Source folder for anchor identity images.
  # > Explicit anatomy close-up subsets are kept separate for sampling control.
  genitals:
    - "6-DICK"   # Source folder for exposed anatomy close-ups.
    - "6-SHEATH" # Source folder for sheathed anatomy close-ups.
  # > Complete body.
  fullbody:
    - "0-FULLBODY" # Full-body identity examples; also included in rest subset.
    - "99-VERSION-LEANER" # Alternate leaner full-body examples.
  # > General body, hand, leg, tail, and leaner variant images.
  body_details:
    - "1-TORSO" # Torso detail examples.
    - "3-HANDS" # Hand detail examples.
    - "4-LEGS" # Leg/foot detail examples.
    - "5-TAIL" # Tail detail examples.
  # > Focused fur-texture crops derived from high-resolution Rook sources.
  fur:
    - "7-FUR" # Fur crops for coat, mane, spots, paws, tail, and anatomy-adjacent texture.
  head:
    - "2-HEAD" # Face/head detail examples.

# > Also export manifest rows marked sfw into an additional generated sfw subset.
export_sfw_subset: true # < Add generated sfw subset from rows marked sfw.

# ==============================================================
# == Images
# ==============================================================

# Export-time mirroring writes fixed extra image/caption pairs before
# SimpleTuner sees the dataset. flip_only keeps kneifftools from applying its
# own random crop, rotation, or color jitter.
augmentations:
  mirrored_extra: true           # Add one mirrored copy per selected source image.
  mirrored_transform: flip_only  # Exact horizontal mirror, no kneifftools crop/jitter.
  seed: 12345                    # Makes mirrored export naming deterministic.

# Export image area bounds before SimpleTuner sees the dataset. The upper bound
# preserves more source detail than Chroma while staying below roughly 1280x1280.
image_resize:
  min_pixel_area: 512   # Upscale tiny images below roughly 512x512 area.
  max_pixel_area: 1280  # Downscale large images above roughly 1280x1280 area.

# ==============================================================
# == Captions
# ==============================================================

# Controlled caption rendering context used by all generated sidecars.
caption:
  subject_sex: male           # Adds male subject wording to prose renderers.
  add_domains_to_tags: false  # Flux2 captions are natural prose, not domain-tag tails.

# Flux2 Klein uses a Qwen-style text stack; use Kneiff's natural-language profile.
caption_outputs:
  mode: hybrid_txt        # One .txt sidecar per image containing selected formats.
  formats: [nlg]          # Natural-language caption profile for Flux2 Klein training.
  tag_scope: supplemental # Kept explicit for consistency with shipped overlays.

# Optional per-subset caption output overrides. Keep commented unless a subset
# needs different caption files or formats.
# caption_outputs_overrides:
#   felkin_anchor:
#     mode: hybrid_txt
#     formats: [nlg]
#   genitals:
#     mode: hybrid_txt
#     formats: [nlg]
#   fullbody:
#     mode: hybrid_txt
#     formats: [nlg]
#   body_details:
#     mode: hybrid_txt
#     formats: [nlg]
#   fur:
#     mode: hybrid_txt
#     formats: [nlg]
#   head:
#     mode: hybrid_txt
#     formats: [nlg]
#   sfw:
#     mode: hybrid_txt
#     formats: [nlg]

# ==============================================================
# == Publishing
# ==============================================================

publishing:
  huggingface:
    pretty_name: Ladybug Felkin # Human-readable dataset-card base title.
    version: v5.6-f2k # Public dataset version label.
    optimized_for_model: FLUX.2 Klein Base 9B # Rendered as [For FLUX.2 Klein Base 9B].
    tags: [flux2, flux2-klein, flux2-klein-9b, flux2-klein-base-9b, lora, diffusion-training, image-captioning]
    adult_content: true # Render the adult-content notice in README.md.

# ==============================================================
# == LoRA Training
# ==============================================================

training:
  enabled: true        # Enables kneifftools training artifact generation.
  backend: simpletuner # Selects the SimpleTuner LoRA backend.

  simpletuner:

    # enabled: true # Optional backend-specific switch; defaults to enabled.

    # Subset-specific SimpleTuner dataloader overrides. Any key here is merged
    # into the subset backend entry after the dataset defaults below.
    subsets:
      felkin_anchor:
        probability: 1          # Sample anchor identity images at steady exposure.
        resolution: 1280        # High-resolution identity anchors support 1280 area.
        # minimum_image_size: 1024 # Avoid using weakly upscaled anchors.
        crop: false             # Preserve full source composition.
        # repeats: 0            # Optional per-subset repeat count.
        # disabled: false       # Optional per-subset on/off switch.
      fullbody:
        probability: 8.0       # Sample full-body subset at high probability.
        resolution: 1280        # Full-body images are the strongest high-res subset.
        crop: false             # Preserve full-body composition.
        # repeats: 0
        # disabled: false
      body_details:
        probability: 3          # Sample body detail subset at normal probability.
        crop: false             # Preserve detail framing; many samples are already close crops.
        # repeats: 0
        # disabled: false
      fur:
        probability: 2.0        # Give focused fur crops extra exposure.
        crop: false             # Preserve texture crop intent; do not random-crop fur samples.
        # repeats: 0
        # disabled: false
      head:
        probability: 2        # Sample head/face details at extra exposure.
        resolution: 1280        # Preserve detail in high-res head crops.
        crop: false             # Preserve detail framing; many samples are close crops.
      genitals:
        probability: 4          # Sample anatomy subset at normal probability.
        crop: false             # Preserve anatomy crop framing.
        # repeats: 0
        # disabled: false

    # Dataset backend defaults copied into every image subset entry.
    dataset:
      type: local                  # SimpleTuner local filesystem backend.
      dataset_type: image          # Image dataset backend, not text embeds.
      resolution: 1024             # Default target area edge for mixed-resolution subsets.
      minimum_image_size: 512      # Keep smaller detail/anatomy crops after export upscaling.
      maximum_image_size: 1280     # Upper bound for input image sizing.
      target_downsample_size: 1280 # Downsample target for cache generation.
      resolution_type: pixel_area  # Area-based buckets, not smaller-edge resize.
      caption_strategy: textfile   # Read captions from exported .txt sidecars.
      metadata_backend: discovery  # Discover images/captions from directories.
      crop: false                  # Preserve source framing for this Flux2 Klein run.
      crop_style: center           # Inert while crop is false; avoids random-crop defaults.
      crop_aspect: preserve        # Preserve original aspect instead of square crop.
      # repeats: 0                 # Dataset-level repeat count; subset can override.
      # disabled: false            # Dataset-level backend switch.

    # Optional shared text-embedding backend overrides. Commented keys show the
    # kneifftools-owned defaults for the generated text_embeds backend.
    # text_embeds:
    #   id: text-embeds        # Backend id used by SimpleTuner.
    #   type: local            # Store embeddings on local filesystem.
    #   dataset_type: text_embeds
    #   default: true          # Marks this as the shared embedding backend.
    #   cache_dir: .simpletuner-cache/text # Override generated text cache dir.
    #   disabled: false

    # Generate a SimpleTuner user_prompt_library JSON from fixed Flux2 prompts
    # plus manifest-derived controls.
    validation_prompts:
      enabled: true
      # positive_prefix: "masterpiece, best quality, score_7, safe, "
      custom:
        f2k_portrait: "Rook_Kaefer is shown in a portrait as a male Felkin character with an expressive face, soft studio lighting, red-and-black ladybug fur markings."
        f2k_fullbody_reference: "Rook_Kaefer is shown from the rear in a full-body view, walking away from the viewer with his back, legs, tail, and red-and-black fur markings clearly visible."
        f2k_body_detail: "Rook_Kaefer is walking sideways across the frame in a clear profile view, with his torso, arms, legs, tail, and fully sheathed genital."
      from_manifest:
        enabled: true
        seed: 42
        profile: nlg

    # Model and LoRA settings merged into the top-level SimpleTuner config.
    model:
      model_type: lora         # Train a LoRA adapter.
      model_family: flux2      # Use SimpleTuner's FLUX.2 model family.
      model_flavour: klein-9b  # Balanced 32GB+ run for the undistilled 9B base.
      lora_rank: 64            # LoRA rank/capacity.
      lora_alpha: 64           # LoRA alpha scaling.
      lora_type: standard      # Standard LoRA, not an alternate adapter type.
      lora_format: comfyui     # Ask SimpleTuner to save ComfyUI-style LoRA keys.
      flux_lora_target: all    # Train all Flux2 LoRA target modules.

      # Base FLUX.2 Klein asset. Klein models use the bundled Qwen3 text
      # encoder, so no separate T5/text-encoder path is set here.
      pretrained_model_name_or_path: black-forest-labs/FLUX.2-klein-base-9B
      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

      # Precision defaults for FLUX.2 Klein LoRA training.
      base_model_precision: no_change
      base_model_default_dtype: bf16
      mixed_precision: bf16
      # init_lora: /path/to/existing.safetensors # Optional LoRA resume/init weights.
      # lycoris_config: /path/to/lycoris.json    # Optional LyCORIS config path.

    # == Curriculum
    curriculum:
      enabled: true
      phases:
        - name: focused_start
          start_step: 0
          subsets: [genitals, fullbody, head]
        - name: full_mix
          start_step: 800 # !! From this step onward, all subsets are active.
          subsets: all

    # Full-run trainer config. trainer_testrun below overrides selected keys
    # only when knf lora sync/train is called with --testrun.
    trainer:
      tracker_project_name: kneifftools-lora-training # Experiment tracker project.
      tracker_run_name: ladybird-f2k9b                # Full-run tracker name.
      hub_model_id: ladybird-f2k9b-lora               # Optional Hub model id if push_to_hub is enabled.

      optimizer: adamw_bf16        # BF16 AdamW optimizer.
      learning_rate: 0.00008       # 1.2e-4; carried over from the earlier F2K run.
      lr_scheduler: constant_with_warmup # Ramp up LR to avoid weird color biases.
      lr_warmup_steps: 100
      seed: 42                     # Training seed.
      gradient_checkpointing: true # Reduce VRAM by recomputing activations.
      max_grad_norm: 1.0           # Clip gradients for stability.
      grad_clip_method: norm       # Clip by gradient norm.

      train_batch_size: 2            # 32GB+ balanced per-device batch size.
      gradient_accumulation_steps: 1 # Keep effective batch at 4 for 9B.
      num_train_epochs: 0            # Step-based training; max_train_steps controls run length.
      max_train_steps: 3000          # Full Flux2 Klein character run.
      ignore_final_epochs: true      # This fixes early stop due to weird epoch calculation by simpletuner

      caption_dropout_probability: 0.0  # Disable caption dropout; keep trigger behavior stable.
      data_backend_sampling: auto-weighting  # Auto-balance dataloader sampling based on subset probabilities and sizes.

      # push_to_hub: false             # Upload final artifacts to Hugging Face Hub.
      # push_checkpoints_to_hub: false # Upload intermediate checkpoints.
      checkpoints_total_limit: 5       # Keep latest 5 checkpoints.
      checkpoint_step_interval: 200    # Save checkpoint every 200 steps.
      validation_step_interval: 200    # Run validation every 200 steps.
      disable_benchmark: false      # Render baseline validation before training.
      validation_prompt_library: false # Disable SimpleTuner built-in prompt library.
      validation_disable_unconditional: true # Skip unconditional validation branch.
      validation_resolution: 1024x1024       # Keep validation affordable for the Flux2 run.
      validation_num_inference_steps: 40     # Full-run validation sampling steps.
      validation_guidance: 5.0               # Inference guidance used by validation.
      validation_guidance_real: null         # Remove Chroma default guidance-real from generated config.
      validation_guidance_rescale: 0.0       # Guidance rescale; 0 disables rescale.
      validation_seed: 42                    # Deterministic validation seed.
      num_eval_images: 1                     # Images per validation prompt.
      flux_guidance_mode: constant           # Keep Flux2 validation guidance stable across prompts.
      flux_guidance_value: 1.0               # Flux2 Klein guidance value from packaged template.

      report_to: tensorboard # Write TensorBoard logs.
      logging_dir: logs      # TensorBoard log dir relative to TRAINING/<config-id>_<run>.

      # Additional kneifftools-owned SimpleTuner defaults. Leave commented to
      # use defaults from src/kneiff/training/lora/simpletuner.py.
      # use_ema: false
      # vae_batch_size: 1
      # data_backend_sampling: uniform
      # allow_dataset_oversubscription: true
      # aspect_bucket_alignment: 64
      tokenizer_max_length: 512
      # t5_padding: zero
      # fully_unload_text_encoder: true
      # offload_during_startup: true
      # keep_vae_loaded: false
      # resume_from_checkpoint: /path/to/checkpoint

    # Short smoke-run overrides. These merge last only with --testrun.
    trainer_testrun:
      disable_benchmark: true
      tracker_run_name: ladybird-f2k9b-testrun # Separate tracker/log name for smoke runs.
      max_train_steps: 12                     # Tiny run to verify plumbing.
      checkpoint_step_interval: 12            # Save at the end of the smoke run.
      validation_step_interval: 12            # Validate at the end of the smoke run.
      validation_num_inference_steps: 12      # Faster validation for smoke runs.