consistent-pet / README.md
bensblueprints's picture
Add files using upload-large-folder tool
d2848ca verified
|
Raw History Blame Contribute Delete
10.7 kB
metadata
language:
  - en
tags:
  - synthetic
  - pet-portraits
  - image-to-image
  - reference-images
pretty_name: Consistent Pet Portrait References
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - metadata.jsonl
          - references/**/*.png

Consistent Pet — portrait reference library

This project uses a 96 GB Blackwell GPU and image-editing models to preserve a pet's identity across portrait references. This release packages the existing Higgsfield template library; it does not enable bulk generation on the live store.

725 distinct reference concepts, all full-resolution PNGs at1792×2240, plus100 separately labeled vintage poster versions. Every master was decoded and SHA-256 matched against its source manifest. There are725 optimized browsing thumbnails. No customer originals, customer portraits, credentials, provider account URLs or private storage keys are included.

  • Hugging Face dataset
  • GitHub repository
  • Open index.html locally for searchable previews and sex/category/collection filters.
  • catalog.json, catalog.csv and ImageFolder-compatible metadata.jsonl provide machine-readable labels, filenames, provenance, prompts, hashes and dimensions.

Collections

Collection Concepts
Original costume references 25
Surreal superheroes, princesses and sports 100
Vintage newspapers, wanted posters and keepsakes 100
Mobsters 100
Superheroes 100
New300 costume collection 300
Total 725

Sex and category organization

Files are organized as references/{male,female,unisex}/{category}/{stable-id}.png. Thumbnails mirror this under previews/; vintage poster versions mirror it under print-variants/.

Male: 62 · Female: 90 · Unisex: 573.

Sex is costume eligibility, not a biological-sex inference from an animal image. Existing intended boy/girl collections retain their labels. Most profession, adventure and vintage designs are unisex. The vintage batch had provisional alternating boy/girl assignments: those were normalized to unisex except explicitly gendered titles/outfits. source_sex and classification_basis preserve the distinction. For a male or female pet, include matching-sex and unisex references.

Category Male Female Unisex
Armor and Tech 0 0 10
Around the World 0 0 20
BatDog 0 0 10
Bonnie and Clyde 1 4 5
British Crime Club 1 2 7
Casino Royals 0 2 8
Cinema posters 1 0 9
Coastal Living 0 0 20
Cosmic Guardians 0 0 10
Country Life 0 0 20
Detectives and Secret Agents 0 0 2
Dogs With Day Jobs 0 0 20
Elemental Heroes 0 0 10
Exploration 0 0 1
Extreme sports 10 10 0
Fantasy and Magic 0 0 4
Foodie Favorites 0 0 20
Funny Dogs Being Dogs 0 0 20
Heists and Getaways 0 0 10
Heroine Collection 0 10 0
Hobbies With Personality 0 0 20
Holidays and Celebrations 0 1 19
Household documents 0 0 10
Keepsakes 0 0 10
Knights and Warriors 0 0 3
Magazine covers 0 0 10
Memorial and Remembrance 0 0 1
Miami Crime Drama 1 2 7
Music and Performance 0 0 1
Music and the Spotlight 0 0 20
Newspapers 0 0 10
Noir City 0 2 8
Original Champions 1 1 8
Outdoor Adventures 0 0 20
Pirates and Naval 0 0 2
Playful Crime Legends 0 0 10
Princess 0 10 0
Retro Futures and Original Characters 0 0 20
Retro advertisements 0 0 10
Roaring Twenties 1 3 6
Royal fantasy 10 0 0
Royalty 1 2 1
Science Fiction 1 0 2
Science and Inventors 0 0 1
Snow-Day Personalities 1 1 18
Spider Dawg 0 0 10
Sports 10 10 0
Sports collectibles 0 0 10
Steampunk and Time Travel 0 0 2
Storybook Adventures 0 0 20
Street Heroes 0 0 10
Studio portraits 1 3 6
Super Dog 0 0 10
Superhero 10 10 0
Surreal adventure 10 10 0
The Pawfather 1 2 7
Underworld Elegance 1 2 7
Up in the Air 0 0 20
Wanted posters 0 0 10
Western 0 0 1
Wheels and Racing 0 0 20
Wonder and Lightning 0 3 7
Working dogs 0 0 10

Download

GitHub stores full originals and thumbnails in Git LFS. GitHub's source ZIP may contain LFS pointers rather than images; clone with LFS for the complete library:

git lfs install
git clone https://github.com/bensblueprints/consistent-pet.git
cd consistent-pet
git lfs pull
cd reference-library
python tools/validate_library.py

Or download from Hugging Face without Git LFS:

pip install huggingface_hub pillow
hf download bensblueprints/consistent-pet --repo-type dataset --local-dir consistent-pet

Metadata follows Hugging Face ImageFolder conventions. The dataset's default image split includes the725 master references; thumbnails and poster variants are not additional training examples. Dataset-viewer indexing is asynchronous.

Quality and reuse notes

These are synthetic concept references, not identity-transfer proofs or inpainting masks. The first surreal female collection has inherited canine-anatomy concerns and is flagged anatomy-review-required; keep those out of automatic customer production until individually corrected/reviewed. Some vintage compositions deliberately contain multiple pets or paired ages; they are flagged separately. The original Dragon Guardian has a secondary creature. The archive retains all requested concepts rather than silently discarding these items.

reference_used records whether generation used an input reference. Retried standalone generations retain their actual effective prompts. Source masters are unmodified; vintage print_variant files include the finished caption layout while file_name is the clean artwork layer. customer_identity_transfer_validated=false is deliberate. Classification does not certify production readiness. No new license grant is made by this organizational upload.

Rebuild and verify

tools/build_library.py accepts the four existing completed-manifest collection directories and the original25 manifest/directory. It copies only explicitly listed image files after hash/decode checks, creates previews and sanitized metadata, and refuses a mismatched inventory. It never scans customer storage or publishes generation/account journals. Then run tools/build_catalog_pages.py and tools/validate_library.py.

Uploads to Hugging Face use tools/upload_huggingface.py with your locally authenticated token. Credentials are never stored in this repository.