Conditional Diffusion Models: DDPM and DDIM

Install Packages

With Python3.8, run the following command to install all the packages listed in the requirements.txt:

pip3 install -r requirements.txt

DDPM (Denoising Diffusion Probabilistic Models)

• All DDPM implementation details are in p1_model.py

• A modified UNet that supports conditional learning by adding time embedding (time steps in the diffusion process) and context embedding (labels)

Train DDPM on MNIST-M / SVHN Dataset

python3 p1_train.py

Generate Digit Images

Use the model you just trained OR download the pretrained model directly:

gdown 1uX845gl81kgc-lHqo8SS78DBz8n1DWmS -O combined_ddpm.pth

Generate images for each digit (0-9) from MNIST-M / SVHN dataset:

python3 p1_inference.py --output_image_dir <your_output_dir> --model_path <your_ddpm_model_path>

For example:

MNIST-M	SVHN

Evaluation

The output images can be evaluated by digit_classifier.py

python3 digit_classifier.py --folder <your_output_dir> --checkpoint <mnistm_ckpt_path>

	MNIST-M	SVHN	Average
Acuracy	99.80%	99.20%	99.50%

DDIM (Denoising Diffusion Implicit Models)

• All DDPM implementation details are in p2_model.py

Generate Human Face Images

python3 p2_inference.py --noise_folder <input_noise_dir> --output_folder <your_output_dir> --unet_model_path <your_unet_path>

With different ETA (Avg. MSE: 1.56):

Interpolate Two Noises

python3 p2_interpolate.py

• Linear interpolation (LERP): Simply linearly mixes the two noise vectors without considering their position on a hypersphere (by setting SLERP=False)
• Spherical Linear Interpolation (SLERP): Along the surface of a hypersphere (by setting SLERP=True)

LERP	SLERP