The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 81, in _split_generators
first_examples = list(islice(pipeline, self.NUM_EXAMPLES_FOR_FEATURES_INFERENCE))
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 32, in _get_pipeline_from_tar
fs: fsspec.AbstractFileSystem = fsspec.filesystem("memory")
~~~~~~~~~~~~~~~~~^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 302, in filesystem
cls = get_filesystem_class(protocol)
File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 239, in get_filesystem_class
raise ValueError(f"Protocol not known: {protocol}")
ValueError: Protocol not known: memory
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 66, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Anima Style Embedding Full-Frame and Face Corpus
Dataset summary
The corpus contains 500,000 person-bearing illustrations for open-set anime style-embedding research. Human artwork and Anima-generated images use independent style identities because an artist's original work and Anima's response to the corresponding artist tag are not the same visual distribution.
| Source directory | Style identities | Images per identity | Images |
|---|---|---|---|
synthetic/ |
5,000 | 50 | 250,000 |
human/ |
5,000 | 50 | 250,000 |
Each sample contains a normalized full-frame image, an anime-face crop when a face is detected, and JSON metadata. Missing detections retain the full-frame image and set face_detection.status to not_detected; they do not receive a fabricated center crop.
Identity and split contract
Each source has 4,000 training identities, 500 unseen validation identities, and 500 sealed unseen test identities. The source prefix is part of the identity:
synthetic:<artist-tag>identifies Anima's response to an artist tag.human:<artist-tag>identifies a selected period of a human artist's work.
Names shared by both sources are not positive pairs. Every identity has 50 images. The synthetic records use 40 training content cells and ten held-out content cells. Human splits are artist-disjoint.
File format
The source directories contain uncompressed WebDataset tar shards. A sample key groups the following members:
<key>.full.webp
<key>.face.webp # present only when a face is detected
<key>.json
The JSON record contains:
| Field | Meaning |
|---|---|
record_id |
Unique key within the combined corpus |
source |
synthetic or human |
style_id |
Source-qualified style identity |
split |
train, validation, or test |
full_image |
Full-frame filename, dimensions, format, and SHA-256 |
face_image |
Face filename, dimensions, format, and SHA-256, or null |
face_detection |
All detections, selected box, detector hash, threshold, padding, and selection rule |
Source-specific generation or Danbooru metadata remains in the same JSON object. Synthetic metadata includes prompt, negative prompt, diffusion seed, SPEED seed, generation configuration, batch provenance, and output mapping. Human metadata includes Danbooru post ID, artist tag, rating, dimensions, source hash, tags, and temporal-screening fields.
Image processing
Full-frame and face images use WebP quality 90. Synthetic PNG outputs are converted to RGB WebP. Human full-frame WebP bytes are retained from the verified ingest shards. The primary face maximizes detector confidence multiplied by bounding-box area. Its square crop uses 25% padding on each side and is resized to 512×512 pixels.
Face detection uses Fuyucchi/yolov8_animeface. Every sample records the exact detector weight SHA-256. Output shards are written through .partial.tar files and renamed only after the shard closes. Final verification checks unique keys, paired members, stored SHA-256 values, source labels, sample counts, split counts, and a regular decode sample.
Synthetic source
Synthetic images were generated with Anima 1.0 Base and Anima Turbo LoRA v0.2 at strength 0.8. Generation uses 768×768 resolution, 11 Euler steps, CFG 1, heterogeneous batch 32 per GPU, SageAttention2, Anima BlockCompile, and Anima SPEED scales [0.5,1.0]. The 50 fixed content cells cover female and male subjects, age groups, ratings, framing, clothing, activities, and backgrounds. Each cell has a distinct seed shared across style identities for controlled comparisons.
Human source
Human records were selected from the public Shio-Koube/Danbooru-2026-parquet-metadata index. Selection requires one artist tag, solo, exactly one of 1girl or 1boy, a short side of at least 512 pixels, an active post, and no AI-generated, collaboration, comic, character-sheet, reference-sheet, multi-view, or text-focus marker. Downloaded source bytes are checked against the Danbooru MD5 before conversion.
The current 24-month density window uses Danbooru created_at quarters as a screening proxy. Danbooru upload time is not evidence of original publication time. Research claims about an artist's period require an independent publication-time audit.
Intended use
The corpus supports open-set style retrieval, prototype learning, full-frame and face-aware style embedding, source-aware representation analysis, and reference-image-to-Anima-condition research. Artist recognition, authorship attribution, identity recognition, and biometric use are outside the research contract.
Content and rights
The corpus includes general, sensitive, questionable, and explicit-rated records. Access requires an environment appropriate for adult material. Synthetic child and teenage prompt cells are clothed and nonsexual.
Human artworks remain copyrighted by their respective creators. The metadata and files do not grant a new license to reproduce or redistribute those works. The repository remains private pending a rights, takedown, and release review. Researchers must honor source removals and applicable platform terms before any public release.
Verification artifacts
metadata/synthetic_verification.json and metadata/human_verification.json contain final image, face, shard, split, byte, and detection-rate counts. metadata/ also preserves both source manifest summaries. A corpus is complete only when both verification files report status: pass with 250,000 samples each.
- Downloads last month
- 180