Conditional-Diffusion-models

This repository implement conditional diffusion model from scratch and train it on the MNIST-M dataset. Given conditional labels 0-9, and generate the corresponding digit images.

MNIST-M Dataset

Generated from MNIST
# of data: 44,800 / 11,200 (training/validation)
# of classes: 10 (0~9)
A subset of MNIST - The digit images are normalized (and centered) in size 28 * 28 * 3 pixels

Model architecture

Training

python train.py --data_root $data_directory

Download checkpoint

checkpoint link

bash download.sh

Sampling Images

python sampling.py --out_dir $output_directory --checkpoint $checkpoint_directory

Fig 1: (0-9) generate images

(a) t=0    (b) t=80    (c) t=160    (d) t=240    (e) t=320    (f) t=400

Figure 2: First ’0’ in different time steps

Reference

Ho, Jonathan, Ajay Jain, and Pieter Abbeel. "Denoising diffusion probabilistic models."
dome272/Diffusion-Models-pytorch