Jannchie/silva-aesthetic

SILVA — Personal Aesthetic Head

▶ Try it in your browser — upload an illustration, see this person's score live.

Scores an illustration by one specific person's taste — not a universal quality model, so it won't match anyone else's preferences. Output is a single number in [0, 1]; higher means more to this person's liking.

Only the head ships here (~7 MB), not an image model. It runs on top of the frozen google/siglip2-so400m-patch14-384 backbone, which silva[backbone] installs and loads for you.

Quickstart

# pip install "silva-scorer[backbone] @ git+https://github.com/Jannchie/silva"
from silva import SilvaScorer

scorer = SilvaScorer.from_pretrained("Jannchie/silva-aesthetic")
print(scorer.score("your_image.jpg"))     # 0.73
print(scorer.score(["a.jpg", "b.jpg"]))    # [0.73, 0.41]

Already have google/siglip2-so400m-patch14-384 embeddings? Skip the backbone and score them directly:

# pip install "silva-scorer @ git+https://github.com/Jannchie/silva"
from silva import EmbeddingAestheticModel

head = EmbeddingAestheticModel.from_pretrained("Jannchie/silva-aesthetic").eval()
score = head(embedding)["calibrated_score"]  # calibrated to the label distribution; ["score"] for raw. embedding: [B, 1152] pooler_output

Scores (held-out test split)

Spearman	Pearson	MAE (1–5)	Top-5%
0.7342	0.7541	0.4141	0.3523

Architecture: embedding[1152] → LayerNorm → MLP [1024, 512, 256] → ordinal head. Trained on one person's private 1-5 ratings, pairwise preferences, and listwise rankings; labels and images not released. Source

Citation

@software{pan2026silva,
  author  = {Pan, Jianqi},
  title   = {{SILVA}: {SigLIP}-based Illustration Visual Aesthetic Scorer},
  year    = {2026},
  url     = {https://github.com/Jannchie/silva},
}

Model version v3.0.0

Released 2026-09-08. Model versions are independent of the Python package version. This is the latest completed non-augmented joint-training baseline (seed 43, dropout 0.5, pairwise/listwise weights 0.05, selected epoch 32). Variant replacement remains disabled because the ablation did not establish improvement. Validation Spearman is 0.721608; the table above reports held-out test metrics of the raw ordinal score, before distribution calibration. Calibration uses the frozen grade manifests, including their test rows, to reproduce the historical label distribution; calibrated scores have no independent test metric here. The small, actively sampled preference validation pool does not establish a general fine-ranking improvement. Training data remain private.

Pin the head to this release:

from silva import EmbeddingAestheticModel, SilvaScorer

head = EmbeddingAestheticModel.from_pretrained("Jannchie/silva-aesthetic", revision="v3.0.0").eval()
scorer = SilvaScorer(head)

Pictoria backfill uses this exact head's calibrated_score, with no library-wide remapping. The previous main revision is preserved as pre-v3-20260908. See release.json for checkpoint hashes and evaluation provenance.