This project explores the application of diffusion models on the MNIST dataset. We conducted a series of experiments to evaluate the impact of various parameters, including noise schedules, diffusion time steps, and sampling methods, on the model's performance. This README provides an overview of the experimental setup, methodology, and key findings.

1. Noise Schedules: Tested four types - linear, cosine, quadratic, and sigmoid.
2. Diffusion Time Steps & beta_end:
   • 200 steps with a beta_end of 0.08
   • 1000 steps with a beta_end of 0.02
3. Sampling Methods: Evaluated both DDPM and DDIM methods.

• Batch Size: 128
• Number of Epochs: 20
• Beta Start: 0.0001

• Fréchet Inception Distance (FID): Used to assess model performance, with lower scores indicating better performance.

In total, 16 experiments were conducted. The best FID score over the 20 epochs for each configuration was selected to represent the outcome of that setup.

• Quadratic: Lowest average FID score.
• Cosine: Best single FID score.
• Linear and Sigmoid: Higher FID scores compared to quadratic and cosine.

• Models with 200 diffusion time steps (beta_end = 0.08) outperformed those with 1000 steps (beta_end = 0.02).
• DDPM performed better with the (time step = 200, beta_end = 0.08) configuration, showing significant improvement.

• DDPM: Outperformed DDIM in terms of average and best FID scores, despite being slower.

• Time Steps: 200
• Beta End: 0.08
• Beta Start: 0.0001
• Noise Schedule: Cosine
• Sampling Method: DDPM

The choice of noise schedule and diffusion steps significantly impacts image quality in diffusion models. The cosine noise schedule and 200 diffusion steps were found to be the most effective, with the DDPM sampling method outperforming DDIM. Careful tuning of these parameters is crucial for optimizing diffusion models for high-fidelity image generation tasks.

• Python 3.8+
• Dependencies specified in environment.yml

1. Clone the repository:

   git clone https://github.com/yourusername/diffusion-models-mnist.git
   cd diffusion-models-mnist

2. Create and activate the conda environment:

   conda env create -f environment.yml
   conda activate diffusion-env

To run the experiments, execute the main.py script:

python main.py

The script will log the training process and save the results in training.log and grid_search_results.json.

• main.py: Script to run the diffusion model experiments.
• environment.yml: Conda environment file with dependencies.
• training.log: Log file containing training details.
• grid_search_results.json: JSON file with the results of the grid search experiments.