zalandoresearch/fashion-mnist

Comprises 70,000 28x28 grayscale images of Zalando clothing across 10 product categories, maintaining byte-for-byte compatibility with the original MNIST format for seamless algorithm migration. Integrated natively into TensorFlow, PyTorch, Keras, and 10+ other ML frameworks, eliminating custom data loading while presenting a more realistic computer vision task than handwritten digits. Addresses MNIST's saturation as a benchmark—modern CNNs achieve 99.7% on MNIST but face genuine classification challenges with Fashion-MNIST's subtle inter-class variations.

12,667 stars. No commits in the last 6 months.