language:
- en
license: cc
size_categories:
- 100K<n<1M
dataset_info:
features:
- name: image
dtype: image
- name: text
dtype: string
splits:
- name: train
num_bytes: 39858266
num_examples: 140000
download_size: 37136812
dataset_size: 39858266
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
MNIST for Diffusion
Training a diffusion model from scratch is pretty cool, why not do so with the canonical "hello world" dataset of computer vision? This dataset matches the sample dataset from this text_to_image.py diffusion tutorial. Specifying ckg/mnist-for-diffusion ought get you off to the races.
This dataset contains two copies of the original MNIST train & test sets. The first half of the dataset contains MNIST images with the string-ified class id (i.e: "1") and the second half has the class id mapped to a natural language name (i.e: "one"). This little data augmentation doubles the number of samples and should result in interesting behavior if you train a U-Net from scratch whilst using a frozen, pre-trained text-encoder!
Thank you LeCun & Cortes for making this dataset available.