silva-scorer

Scores an illustration by one person's aesthetic taste — an ordinal-regression head on top of frozen google/siglip2-so400m-patch14-384 embeddings. Only the head ships (~7 MB); it is not a universal quality model and won't match anyone else's preferences.

Install

The PyPI name is silva-scorer; it imports as silva.

pip install silva-scorer              # embedding -> score (torch + huggingface-hub only)
pip install "silva-scorer[backbone]"  # image -> score, adds the SigLIP2 backbone + `silva` CLI

Score an image

With the [backbone] extra, silva loads SigLIP2 for you — give it a path (or a list of them) and get a score back:

from silva import AestheticScorer

scorer = AestheticScorer.from_pretrained("Jannchie/silva-aesthetic")
scorer.score("image1.jpg")                  # 0.7421
scorer.score(["image1.jpg", "image2.jpg"])  # [0.7421, 0.3128]

Or from the CLI:

silva score image1.jpg image2.jpg --repo-id Jannchie/silva-aesthetic
# image1.jpg  score=0.7421

Score from an embedding

Already running google/siglip2-so400m-patch14-384 yourself? The core install (no transformers) scores a 1152-d embedding directly:

import torch
from silva import HubAestheticModel

head = HubAestheticModel.from_pretrained("Jannchie/silva-aesthetic").eval()
emb = torch.randn(1, 1152)        # raw pooler_output from the backbone above
print(head(emb)["score"].item())  # [0, 1] — fraction of quality bars cleared

The embedding must be the raw pooler_output of that exact backbone — it's what the head was trained against.

Training your own head

silva is inference-only. To fit a head on your own 1–5 ratings, see the silva-train package.

License

MIT