Pro-DDPM: Progressive Growing of Variable Denoising Diffusion Probabilistic Models for Faster Convergence
Rohit Gandikota (Northeastern University),* Nicholas Brown (Northeastern University)
The 33rd British Machine Vision Conference
Abstract
We describe a new training methodology to train Denoising Diffusion Probabilistic Models (DDPM) for faster convergence speeds. DDPMs have achieved high-quality image synthesis in non-adversarial training; however, their training is computationally intensive due to the Markov chain simulation and sampling for many time steps. The key idea of this work is to progressively increase the network depth of the diffusion model sinusoidally, starting from a low resolution and adding layers as they converge. This showed a 3 times improvement in training speeds while ensuring unprecedented image quality content generation. We compare our model's training performance with the original DDPM, the improved DDPM, diffusion implicit models, and generative models in terms of FID and log-likelihood. We analyse the image quality and variance of the synthesised images on CelebA-HQ, Animal Faces-HQ, Imagenet and CIFAR-10 datasets. We also describe several implementation details of our training method that are important for quality preservation and faster convergence.Video
Citation
@inproceedings{Gandikota_2022_BMVC,
author = {Rohit Gandikota and Nicholas Brown},
title = {Pro-DDPM: Progressive Growing of Variable Denoising Diffusion Probabilistic Models for Faster Convergence},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0121.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by