CoolFace
Datasetpublic

siik/mirrai-sd-training-open-components

MirrAI SD Training Open Components This dataset repo contains the redistributable subset that was actually used in the MirrAI Stable Diffusion hair inpainting training pipeline. Included In This Release face_sketches_refined_generation/ images_512/: 139 processed 512px training images masks/: hair, face protect, and cloth protect masks controls/canny/: canny control inputs face_crops/: identity-preserving face crops manifests/recon_hair_external.jsonl: training… See the full description on the dataset page: https://huggingface.co/datasets/siik/mirrai-sd-training-open-components.

sourceHugging Faceapache-2.0updated 6mo agoView on Hugging Face
0likes33downloads
Dataset Card

MirrAI SD Training Open Components

This dataset repo contains the redistributable subset that was actually used in the MirrAI Stable Diffusion hair inpainting training pipeline.

Included In This Release

  • face_sketches_refined_generation/
  • images_512/: 139 processed 512px training images
  • masks/: hair, face protect, and cloth protect masks
  • controls/canny/: canny control inputs
  • face_crops/: identity-preserving face crops
  • manifests/recon_hair_external.jsonl: training manifest for the open subset
  • reports/summary.json: preprocessing summary

How It Was Used

  • This subset was used as the open long-tail external component for stage 3 and stage 4 SD LoRA training.
  • The processed rows come from the project-owned siik/FaceSketches-HairStyle40-refined subset.
  • Accepted rows: 139
  • Split counts: train 136, val 2, test 1
  • Main style families: default 55, alt_asian_longtail 21, garment_reveal_support 21, coily_protective 15, punk_editorial 14, retro_structured 13

Not Included

The full base training data is intentionally excluded from this Hugging Face release.

  • Excluded: dataset_build/processed/celeba_dialog_hq_generation
  • Excluded: mixed long-tail manifests that enumerate CelebA-derived rows

Why this is excluded:

  • The base training set is derived from CelebA-Dialog HQ and CelebAMask-HQ.
  • The official CelebAMask-HQ agreement limits usage to non-commercial research and says portions of the dataset and derived data must not be further copied, published, or distributed except for internal use within the same organization.
  • Because of that, this repo publishes only the redistributable open component and leaves the restricted base set out of the release.

Reproducing The Full Private Training Set

To rebuild the full private training set locally, use the source repository:

  • GitHub: https://github.com/JangWS1030/SKN22-Final-1Team-AI

Relevant scripts:

  • scripts/download_celeba_dialog_hq.py
  • scripts/preprocess_celeba_dialog_generation.py
  • scripts/build_generation_training_subsets.py
  • scripts/preprocess_external_hair_generation.py
  • scripts/build_longtail_training_subsets.py

Source Attribution

  • Open source subset used here: siik/FaceSketches-HairStyle40-refined
  • Upstream source referenced by that subset: yikaiwang/FaceSketches-HairStyle40