You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
1.2 KiB
1.2 KiB
Tiny DDPM (and DDIM)
This is a bare bones and simple DDPM (Denoising Diffusion Probabilistic Models) implementation on PyTorch. The whole implementation (model + training + sampling) does not exceed 400 lines of code. The training setup and U-Net model loosely resemble the description of the original paper, but it is not a 1 to 1 implementation.
These images were generated after training on CIFAR-10 for 256 epochs on a single RTX 4090.
Usage
Installation
It is recommended (but not required) to use uv to replicate the Python environment:
uv sync # If using uv
python -m pip install . # Otherwise
Training
uv run src/simple_ddpm/train.py # If using uv
python src/simple_ddpm/train.py # Otherwise
Sampling
uv run src/simple_ddpm/sample.py # If using uv
python src/simple_ddpm/sample.py # Otherwise
By default it will perform DDPM sampling. If you want to use DDIM, simply
change the function called at the bottom of src/tiny-ddpm/sample.py to call
ddim_sample_images instead of ddpm_sample_images.