Devanagari Characters Image Dataset

Dataset Summary

The Devanagari Characters Image Dataset is a high-resolution dataset designed to support research and experimentation in generative modeling, specifically for the Hindi script. It includes images for:

Vowels (स्वर)
Consonants (व्यंजन)
Matra combinations (e.g., का, कि, की, कु)
Hindi numerals (०-९)

The dataset was created to address the limitations of existing Devanagari datasets, which often suffer from low resolution (typically 32x32 pixels) and limited font variety. This dataset provides a more robust and scalable resource for training advanced generative models such as diffusion models. But it can be used for some other problem statement as well.

Use Cases

Hindi character generation using diffusion models
Full sentence-level generation and rendering
OCR pretraining and benchmarking for Devanagari
Font style transfer and augmentation
High-resolution classification tasks

Dataset Details

Characters Covered: 80+ (including vowels, consonants, matra combinations, and numerals)
Font Variations: 305 unique Devanagari Unicode fonts
Resolution: 128x128 pixels (grayscale)
Images per Character: ~305
Total Size: ~24,000 images

Each character image is rendered programmatically using Unicode-compliant Devanagari fonts to ensure consistency and readability across styles.

Character Coverage

The dataset includes:

Vowels (स्वर)

अ, आ, इ, ई, उ, ऊ, ऋ, ॠ, ऌ, ॡ, ए, ऐ, ओ, औ, अं, अः

Consonants (व्यंजन)

क, ख, ग, घ, ङ, च, छ, ज, झ, ञ, ट, ठ, ड, ढ, ण, त, थ, द, ध, न, प, फ, ब, भ, म, य, र, ल, व, श, ष, स, ह, ळ, क्ष, त्र, ज्ञ

Matra Combinations (e.g.,)

का, कि, की, कु, कू, कृ, के, कै, को, कौ, कं, कः, गा, गि, गु, गो, etc.

Hindi Numerals

०, १, २, ३, ४, ५, ६, ७, ८, ९, १०

Motivation

While exploring the application of diffusion models for multilingual character generation, a key bottleneck was the lack of high-resolution and stylistically diverse datasets in Hindi. This dataset was created to fill that gap. It provides a clean, diverse, and high-resolution dataset suitable for generative tasks and deep learning experimentation.

Limitations

The dataset is synthetically generated using fonts and does not include handwritten data.
Some fonts may render characters with minor stylistic irregularities, especially in matra combinations.