Signatures Dataset

A collection of handwritten signature images with associated names, useful for signature generation and verification research.

Dataset Structure

This dataset contains two splits:

Split	Images	Description
raw	49,530	Original images (various sizes, colors, formats)
processed	48,860	Preprocessed images (512×256, grayscale, white background, black signature)

Columns

Both splits contain:

• image: The signature image
• file_name: Original filename
• label: Numeric label
• name: The person's name (for text conditioning)

How to use it?

from datasets import load_dataset

# Load raw (original) data
raw = load_dataset("ifkash/signatures", split="raw")

# Load processed (preprocessed for training)
processed = load_dataset("ifkash/signatures", split="processed")

# Access a sample
sample = processed[0]
print(sample['name'])   # e.g., "rafael casanova"
sample['image'].show()  # Display the signature

Processed Split Details

The processed split has been preprocessed for machine learning:

• Resolution: 512 × 256 pixels
• Color: Grayscale (L mode)
• Background: White (255)
• Signature: Black (0)
• Aspect ratio: Preserved with white padding

Preprocessing Pipeline

1. Convert to grayscale
2. Invert if dark background detected
3. Contrast stretching
4. Resize with padding to 512×256

Quality Filters Applied

Removed from processed split:

• Black blobs (>60% dark pixels)
• Gray backgrounds (corner mean < 200)
• Nearly blank images (<1% ink)
• Corrupted/unreadable files

Use Cases

• Text-to-Image: Generate signatures from names using models like FLUX, Stable Diffusion
• Signature Verification: Train classifiers to verify authentic signatures
• Style Transfer: Learn signature styles and apply to new names
• Handwriting Analysis: Study signature characteristics

Citation

If you use this dataset, please cite:

@dataset{signatures_dataset,
  author = {ifkash},
  title = {Signatures Dataset},
  year = {2026},
  url = {https://huggingface.co/datasets/ifkash/signatures}
}