javazeroo / diffusion_distillation

diffusion_distillation

TL;DR:

Simply run

bash run.sh

---

Env

conda env create -f dd.yml

Train

1. Adjusting training profiles: diffusion_distillation/config/mnist_base.py
2. Run: diffusion_distillation.ipynb, The output checkpoint will be saved in /tmp/flax_ckpt/checkpoint/ directory.
3. Copy the checkpoint to a safe folder, as the /tmp folder will be case on system reboot.

Sampling

1. Adjusting sampling profiles: diffusion_distillation/config/mnist_distill.py
2. Run the following code to generate images:

python sample_origin.py --num_imgs 1024 --batchsize 64 --startbatch 0 --db_path data/mnist_origin_debug --ckpt_path /path/to/ckpt

• db_path means the output dataset, which will be used in DSNO-pytorch
• num_imgs means the num of images to be generated