guyshilo12/painting-style-recommender

🎨 Painting Recommender (Embeddings + Similarity)

A visual recommendation system that suggests the Top-3 most similar paintings from a dataset, based on CLIP image embeddings and cosine similarity.

✅ Live Demo

Upload a painting image and click Recommend Top-3 to get similar artworks.

📦 Dataset

Dataset: keremberke/painting-style-classification (Hugging Face) Modality: Images (paintings) Total size: 6,417 images

• —Train: 4,493
• —Validation: 1,295
• —Test: 629

Labels: 27 painting styles (e.g., Cubism, Impressionism, Baroque, Fauvism, etc.) Image format: RGB Resolution (most images): 416×416

Why this dataset?

The dataset includes diverse artistic styles with clear visual differences in color palettes, composition, and abstraction levels — ideal for embedding-based similarity and clustering.

🔎 Exploratory Data Analysis (EDA)

Basic sanity checks and statistics were performed:

• —dataset split distribution
• —number of styles (labels)
• —label distribution (imbalanced but acceptable for similarity-based recommendation)
• —visual samples across styles

🧠 Embeddings

Model: openai/clip-vit-base-patch32 (Hugging Face Transformers) Embedding size: 512-dim vectors

Why CLIP?

CLIP captures high-level visual semantics (style, composition, color patterns), making it suitable for similarity search and recommendation in visual domains.

Embeddings are L2-normalized, so cosine similarity can be computed efficiently using dot product.

🧩 Embeddings Analysis: Dimensionality Reduction + Clustering

To analyze the embedding space we used a balanced subset (up to ~100 images per style) to stay computationally efficient while preserving structure.

Dimensionality reduction: PCA → UMAP (2D visualization) Clustering: K-Means (k=10)

Cluster interpretation

Clusters are generally coherent and group paintings by shared visual properties such as:

• —abstraction level (abstract/cubist vs. classical)
• —dominant color palettes
• —composition/geometry
• —texture/brushstroke patterns

Importantly, clusters do not perfectly replicate labels — which is expected and desirable for a similarity-based recommender.

🎯 Recommendation Method

Input

• —Image Upload (user provides an image)

Output

• —Top-3 recommended paintings from the dataset

Steps

1. 1.Convert user image → CLIP embedding
2. 2.Compute cosine similarity against stored dataset embeddings
3. 3.Retrieve the Top-3 highest similarity scores and return corresponding paintings

Example output includes similarity scores such as:

• —Style: Cubism | Similarity: 0.847
• —Style: Fauvism | Similarity: 0.834
• —Style: Cubism | Similarity: 0.833

💾 Files in this Repo

• —app.py – Gradio application
• —requirements.txt – dependencies
• —painting_embeddings.parquet – saved embeddings for recommendation

🎥 Presentation Video

A 3–5 minute presentation demonstrating:

• —the working application
• —dataset overview
• —embeddings and clustering analysis
• —recommendation logic

📎 Video link: https://drive.google.com/file/d/1TJnNUZmeUHltg9FOC5y1BRXugQL1Fbqi/view?usp=sharing

📓 Development Notebook

The full development process of the project is documented in the following Jupyter Notebook, including:

• —dataset loading and exploratory data analysis (EDA)
• —image embedding generation
• —dimensionality reduction and clustering
• —recommendation logic implementation

📎 Notebook link: https://huggingface.co/spaces/guyshilo12/painting-style-recommender/resolve/main/CopyofAssignment3Embeddings%2CRecSys%2CSpaces.ipynb