ladyFaye1998/artsleuth-weights

ArtSleuth · Hub weights

Production checkpoints for computational art analysis — DINOv2 + CLIP fusion, multi-task heads, and interpretable attributions.

Project documentation · Danielle Lesin on Hugging Face

---

Overview

This repository hosts official PyTorch weights for ArtSleuth: a framework for brushstroke-level analysis, style and genre classification, artist attribution, forgery-oriented screening, and visualization-oriented explanations — built on modern vision backbones.

Resource	Link
Interactive demo	🤗 Space · ArtSleuth
Application source	github.com/ladyFaye1998/ArtSleuth
Training data (reference)	huggan/wikiart

Files

File	Description
style_head.pt	Linear head: CLIP embedding → 27 WikiArt style classes
genre_head.pt	Linear head: CLIP embedding → 11 WikiArt genre classes
artist_head.pt	Linear head: CLIP embedding → 129 WikiArt artist classes
fusion_head.pt	Cross-attention fusion module (DINOv2 + CLIP)
taxonomy.json	Canonical label lists (styles, genres, artists)
best_sota.pt	Full checkpoint: partially unfrozen backbones + all heads (~2 GB)

Training summary

Heads were trained on WikiArt (81,444 images; 27 styles, 129 artists, 11 genres):

• Backbones: DINOv2 ViT-B/14 + CLIP ViT-L/14 (last 3 transformer blocks unfrozen)
• Loss: Multi-task cross-entropy + supervised contrastive (SupCon)
• Optimiser: AdamW, cosine schedule with warmup, mixed precision (fp16)
• Schedule: 5 epochs, effective batch size 64 (gradient accumulation)
• Hardware: Single 16 GB GPU

Benchmarks (WikiArt, 80/20 split, macro-averaged)

Configuration	Style (27)	Artist (129)	Top-5 artist	Genre (11)
Fine-tuned (this checkpoint)	67.1 %	79.0 %	96.9 %	76.6 %

Usage

# Recommended: ArtSleuth pulls small heads from this repo when needed
import artsleuth
result = artsleuth.analyze("painting.jpg")
print(result.summary())

# Manual download (full checkpoint)
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="ladyFaye1998/artsleuth-weights",
    filename="best_sota.pt",
)

Limitations

• School head: Randomly initialised — no labelled school data; requires fine-tuning before use.
• Compute: 5-epoch fine-tune on one GPU; more compute would likely improve metrics.
• Domain: Predominantly Western easel painting (15th–20th century). Other traditions are not yet evaluated.

Citation

@software{lesin2026artsleuth,
  author  = {Lesin, Danielle},
  title   = {{ArtSleuth}: Computational Art Analysis Framework},
  year    = {2026},
  url     = {https://github.com/ladyFaye1998/ArtSleuth},
  license = {MIT}
}

License

MIT — see LICENSE.

---

_{Weights published by Danielle Lesin · Hugging Face · Kaggle}