← Back to projects
Jupyter · PyTorch · 2025

Diffusion Model

A from-scratch study of diffusion models, with experiments varying quantization and the number of timesteps to study their effects.

01

Overview

The project implements and trains different diffusion models from scratch, varying quantization and timestep choices to investigate how those decisions affect the resulting models.

FocusUse a controlled implementation to make the trade-offs between diffusion-model design choices more visible.
02

Experiment space

The repository is organized around comparing models under different configurations, with particular attention to quantization and the number of timesteps used in the diffusion process.

Experiment results and the accompanying project report live in the repository.
03

Implementation

The project uses PyTorch and includes a dedicated pytorch-diffusion implementation directory. The repository is intentionally kept close to the research/engineering workflow: code, data, and a written project report are all tracked together.

This makes it a useful companion to the other ML projects in the portfolio: rather than treating a model as a black box, the implementation is the subject of the project itself.

04

Artifacts

The GitHub repository currently includes the project report and supporting experiment data alongside the implementation.

I’m leaving the quantitative conclusions intentionally light here until there is a richer public README to draw from. The next version of this page can incorporate the report's charts and findings directly.