🌸 DE-PERFUME

Multimodal Synthetic Perfume Dataset (Text + Image)

✨ Overview

DE-PERFUME is a synthetic multimodal dataset of perfume concepts.
Each sample links a natural-language user request to structured attributes and an AI-generated perfume bottle image.

• 📦 Records: 100
• 🧠 Modalities: Text + Image
• 🎯 Primary use case: Text → Image retrieval

---

📂 Dataset Contents

• perfume_dataset.csv — main metadata table
• perfume_images.zip — all generated images
• eda_outputs/ — EDA figures used below

---

🧾 Table Schema

• id — unique numeric identifier
• user_request — natural-language perfume request
• scent — scent family (Citrus, Woody, Floral, etc.)
• style — bottle style (Minimalist, Luxury Gold, Vintage, etc.)
• image_path — image filename inside perfume_images.zip

---

🔗 Image Linking

Each row references its image via the image_path column
(e.g. perfume_23.png, stored inside perfume_images.zip).

---

🧪 Data Generation

Text and metadata were synthetically generated using predefined scent and bottle-style categories.
Images were generated using a diffusion model conditioned on these attributes and a clean product-photography style.

---

📊 Exploratory Data Analysis (EDA)

1️⃣ Prompt Length by Scent × Style

Explanation:
This heatmap shows the average length (in characters) of user requests for each scent–style combination.
The results indicate a highly controlled and consistent text-generation process, with only minor variation across categories.

---

2️⃣ Prompt Length Distribution

Explanation:
The histogram displays the distribution of prompt lengths across the dataset.
Most samples fall within a narrow range, with no extreme outliers, making the dataset well-suited for embedding and retrieval tasks.

---

3️⃣ Mean Image Color per Scent

Explanation:
For each scent category, the mean RGB values of all associated images were computed and visualized.
Subtle differences between scent families suggest that prompt conditioning influenced visual color characteristics.

---

4️⃣ Image Diversity: Brightness vs Contrast

Explanation:
This scatter plot compares image brightness (mean grayscale intensity) with contrast (grayscale standard deviation).
The broad spread of points indicates visual diversity and no evidence of mode collapse in image generation.

---

🎯 Intended Use

This dataset is intended for:

• multimodal retrieval (text ↔ image)
• evaluation of multimodal embeddings (e.g., CLIP-style models)
• educational demonstrations and prototyping

---

⚠️ Limitations

• Fully synthetic data (subject to generative model biases)
• Small-scale dataset (100 samples), designed as a focused prototype

---

📜 License

MIT

---

📌 Citation

@dataset{de_perfume_2026, title={DE-PERFUME: Multimodal Synthetic Perfume Dataset (Text + Image)}, authors={Omry Nadiv, Peleg Elraz}, year={2026}, publisher={Hugging Face} }