Warden Ship — art-domain CLIP
openai/clip-vit-base-patch32 fine-tuned contrastively on (painting, caption)
pairs built from WikiArt metadata, so both encoders speak the vocabulary of art
history. Used by Warden Ship for typed
semantic search and for every image embedding it stores.
Exported as a standard HF CLIP directory — it loads with the same call as the base model.
from transformers import CLIPModel, CLIPProcessor
model = CLIPModel.from_pretrained("breskvarmatej/warden-ship-clip-art")
processor = CLIPProcessor.from_pretrained("breskvarmatej/warden-ship-clip-art")
Zero-shot accuracy on the held-out WikiArt test set
| Model | Style (27 classes) | Genre (10 classes) |
|---|---|---|
openai/clip-vit-base-patch32 | 25.8 % | 57.7 % |
| this model | 55.3 % | 79.8 % |
Caveats
Fine-tuning moved the image-embedding space. It keeps the same 512-d projection as base CLIP, so vectors from the two models are interchangeable in shape but not in meaning — a head fitted on base-CLIP vectors will silently score this model's output in a space it never saw, with no error to warn you. Don't mix them.
Trained on WikiArt, which is heavily weighted toward European painting of the 19th and 20th centuries; accuracy outside that is materially worse.