Warden Ship — ArcFace style-retrieval embedding
A metric embedding trained with ArcFace loss so that works in the same style land near each other. Warden Ship uses it to lay out and cluster its art-history map (F5), where raw CLIP put visually similar but stylistically unrelated works together.
TorchScript. Takes float32 images in [0, 1], shape [B, 3, 224, 224]
(normalisation is baked in), returns L2-normalised 512-d embeddings.
import torch
from huggingface_hub import hf_hub_download
path = hf_hub_download("breskvarmatej/warden-ship-f1-embed", "f1_embed_model.pt")
model = torch.jit.load(path, map_location="cpu").eval()
embeddings = model(images) # [B, 3, 224, 224] in [0, 1] -> [B, 512]
Style retrieval on the held-out test set
| Embedding | precision@1 |
|---|---|
| CLIP ViT-B/32 image features | 0.416 |
| this model | 0.610 |
Evaluated on a 2000-image sample of the held-out split, not the full test set.
Caveats
Trained on WikiArt: European painting of the 19th and 20th centuries dominates. The embedding is tuned for style neighbourhoods — it is not a general-purpose visual descriptor, and it carries no notion of subject or period.