ij/anima-style-factor-intervention-v4

Anima Style Factor Intervention v4

Summary

Anima Style Factor Intervention v4 is a private training feature dataset for person-oriented open-set style transfer. It pairs deterministic Line, Color, Texture, and Geometry interventions with frozen SigLIP2 full/face features and frozen Anima internal descriptors. The dataset is derived from the private 500,000-image Anima style corpus and does not redistribute a second copy of the source images.

The training split contains 104,000 transformed records from 72,000 source artworks. The validation split contains 4,096 transformed records from 512 held-out style identities. Synthetic and human sources are balanced.

Intended use

The records supervise a factorized relational style encoder. Each transformed record is paired with its original source record through source_record_id. Training measures factor, direction, strength, monotonicity, non-target leakage, and style-identity invariance. Transformed copies are intervention pairs only; they must not be counted as independent reference artworks.

Transform families

Factor	Training families	Validation-only family
Line	dilate_erode, blur_sharpen, darkness_contrast	edge_overlay
Color	palette_remap, split_tone, tone_curve	channel_mixer
Texture	smooth_detail, frequency, grain_noise	median_speckle
Geometry	crop_zoom_translate, perspective, lens_warp	shear

Levels are weak, medium, and strong. Signed operations balance positive and negative directions. Lens-warp coefficients are bounded to 0.035, 0.055, and 0.075. Transform version lens-safe-v4 and each record's deterministic operation_seed fully specify image-space preprocessing.

Feature contracts

Feature	Shape per record	Dtype	Source
SigLIP2 full image	30 x 1152	BF16	timm/vit_so400m_patch16_siglip_gap_512.v2_webli
SigLIP2 face crop	30 x 1152	BF16	Same backbone and automatic anime-face crop
Face mask	Scalar	Boolean	Face availability after the transformed crop
Anima internal descriptor	3 x 4096	BF16	Frozen Anima blocks 8, 18, and 26

Anima descriptors use a 768x768 transformed image, sigma 0.1, noise seed 20260715, and the frozen empty-prompt path. Each descriptor concatenates the spatial mean and log standard deviation at the selected block.

Repository layout

README.md
manifest.jsonl
manifest.summary.json
verification/
  siglip_verification.json
  anima_pilot_verification.json
  anima_train_verification.json
  anima_incremental_signal_report.json
full_external/
  intervention-w*-p*.json
  intervention-w*-p*.safetensors
  intervention-worker-*.json
anima_pilot_features/
  anima-w*-p*.json
  anima-w*-p*.safetensors
  anima-worker-*.json
anima_train_features/
  anima-w*-p*.json
  anima-w*-p*.safetensors
  anima-worker-*.json
code/
  factor_interventions.py
  build_factor_intervention_manifest.py
  extract_factor_intervention_features.py
  extract_anima_intervention_features.py
  verify_factor_intervention_cache.py
  verify_anima_intervention_pilot.py
SHA256SUMS

The JSON metadata beside every tensor part contains its ordered record list and preprocessing contract. Consumers must join modalities by record_id, not by filesystem order.

Integrity

The canonical manifest contains 108,096 unique records and has SHA-256:

be1fd38d64b8fe512e0ce5514dc5178d4a0e1e86ce724af8d022f689164a4b8b

Accepted caches must satisfy all of the following:

• 104,000 unique training records and 4,096 unique validation records;
• exact coverage of the selected manifest rows;
• matching ordered IDs in JSON metadata and tensor rows;
• finite BF16 tensors with the documented shapes;
• matching transform lineage and preprocessing metadata;
• matching file digests in SHA256SUMS.

Source data and access

Source images remain in the private ij/anima-style-embedding-500k-full-face dataset. Synthetic images were generated by Anima; human images retain their original creator rights and are provided only for the authorized research workflow. No public redistribution license is granted by this feature repository.

Limitations

The corpus contains person-bearing anime and illustration images. It does not establish factor control for non-person imagery. The four factors are operational categories defined by controlled transformations, not a claim that artistic style has a unique or perfectly independent decomposition.