README.md

metadata

library_name: keras
tags:
  - generative
  - denoising
  - diffusion
  - ddim
  - ddpm
  - unconditional-image-generation

This model was created for the Keras code example on denoising diffusion implicit models (DDIM).

Model description

The model uses a U-Net with identical input and output dimensions. It progressively downsamples and upsamples its input image, adding skip connections between layers having the same resolution. The architecture is a simplified version of the architecture of DDPM. It consists of convolutional residual blocks and lacks attention layers. The network takes two inputs, the noisy images and the variances of their noise components, which it encodes using sinusoidal embeddings.

Intended uses & limitations

The model is intended for educational purposes, as a simple example of denoising diffusion generative models. It has modest compute requirements with reasonable natural image generation performance.

Training and evaluation data

The model is trained on the Oxford Flowers 102 dataset for generating images, which is a diverse natural dataset containing around 8,000 images of flowers. Since the official splits are imbalanced (most of the images are contained in the test splite), new random splits were created (80% train, 20% validation) for training the model. Center crops were used for preprocessing.

Training procedure

The model is trained to denoise noisy images, and can generate images by iteratively denoising pure Gaussian noise.

For more details check out the Keras code example, or the companion code repository, with additional features..

Training hyperparameters

Hyperparameters	Value
num epochs	80
dataset repetitions per epoch	5
image resolution	64
min signal rate	0.02
max signal rate	0.95
embedding dimensions	32
embedding max frequency	1000.0
block widths	32, 64, 96, 128
block depth	2
batch size	64
exponential moving average	0.999
optimizer	AdamW
learning rate	1e-3
weight decay	1e-4

Model plot

View model plot