Jannchie/silva-luna

SILVA — Distilled VLM Judge

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

Scores an illustration by the taste of a VLM judge, distilled — 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 judge'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-luna")
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-luna").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%
0.8105	0.7986	0.4331	0.4672

Architecture: embedding[1152] → LayerNorm → MLP [1024, 512, 256] → ordinal head. Trained on rankings from openai:gpt-5.6-luna, which ordered eight illustrations at a time; the orderings were pooled with Plackett-Luce into one latent per picture (degree 14, split-half reliability 0.827). The labels ARE published -- see the dataset link above. Source

Citation

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