Home-Made Diffusion Model

This project demonstrates a diffusion-based generative model that I built and trained entirely from scratch.
The goal was to explore whether it is possible to design and train a functional diffusion model with minimal resources, while still producing images that carry some structure and coherence.

Overview

• Architecture: Fully custom-built diffusion model.
• Dataset: Scraped manually with a custom script (~400 MiB).
• Training: Conducted on Kaggle with limited GPU resources.
• Results: The model generates images at multiple resolutions (128p and 256p) that, while imperfect, display meaningful structure.

Preview samples:

Files

• Full-Project.ipynb – Main notebook containing model architecture, dataset preprocessing, and training pipeline.
• Inference.ipynb – Script for running inference and generating new samples.
• Diffusion-Model-Trained-5h.pth – Model checkpoint trained for 5 hours.
• Diffusion-Model-Trained-10h.pth – Model checkpoint trained for 10 hours.
• preview-grid-1-256p.png, preview-grid-2-128p.png – Generated image grids for visual evaluation.
• LICENSE – Project license.

Usage

1. Clone the repository and open either notebook (Full-Project.ipynb or Inference.ipynb) in Jupyter or Kaggle.
2. Load one of the trained checkpoints (.pth files).
3. Run the inference script to generate new samples.

Motivation

Most diffusion models require massive datasets and compute power.
This project aims to show that, even with constrained resources and a small dataset, it is possible to experiment with generative diffusion processes and obtain meaningful outputs.