DDIM.jl

This repository is now merged into Lux.jl examples.

Lux.jl implementation of Denoising Diffusion Implicit Models (arXiv:2010.02502).

The implementation follows the Keras example.

The model generates images from Gaussian noises by denoising iteratively.

Usage

Install Julia and instantiate Project.toml.

Follwoing scripts are tested on a single NVIDIA Tesla T4 instance.

Dataset

Download and extract Dataset images from 102 Category Flower Dataset.

Training

$julia --project train.jl \
    --dataset-dir oxford_flower_102 \ # path to dataset directory containing image files
    --epochs 25 \
    --image-size 96 \
    --batchsize 64 \
    --learning-rate 1e-3 \
    --weight-decay 1e-4 \
    --val-diffusion-steps 80 \
    --output-dir output/train # path to save checkpoint and images

You can also change model hyper parameters. See main function in train.jl.

Image generation

$julia --project generate.jl \
    --checkpoint output/ckpt/checkpoint_25.bson \ # path to checkpoint
    --image-size 96 \
    --num-images 10 \
    --diffusion-steps 80 \
    --output-dir output/generate # path to save images

You can also change model hyper parameters. See main function in generate.jl. The hyper parameters should be the same during training and generation.