Diffusion Models for Custom Dataset Generation and Augmentation

This repository investigates how well diffusion models (DDPMs) can generate realistic images for a small, labeled dataset of facial images. The primary objective is twofold:

1. Evaluate performance of DDPMs on a personalized dataset.
2. Test their usefulness for data augmentation in downstream classification tasks.

---

Project Structure

File	Purpose
00_image_extraction.ipynb	Prepares and visualizes the custom image dataset of 3 individuals
01_diffusion_model_training.ipynb	Trains a vanilla DDPM on the face images
02_generating_images_from_model.ipynb	Uses the trained DDPM to generate synthetic samples
03_diffusion_model_tester.ipynb	Compares model checkpoints and output quality
05_generating_Images_RECAP.ipynb	Summary notebook of image generation performance
06_diffusion_model_conditional.ipynb	Trains a class-conditional DDPM with label embeddings
07_diffusion_model_conditional.py	Full pipeline training script with Hugging Face accelerate
08_stable_diffusion.py	Fine-tunes Stable Diffusion v2.1 using LoRA for each class prompt
09_Stable_Diffusion_Image_Generation.ipynb	Uses the fine-tuned LoRA model to generate new images
10_Classification_Stable_Diffusion_Augmentation.ipynb	Tests whether adding generated images improves classifier performance
OliverGerardoo_Term_Poster-1.pdf	Summary of applying stable diffusion for image augmentation and comparing which method works the best

---

Goals & Motivation

This project explores:

• How well diffusion models can learn from a limited, real-world dataset.
• Whether generated samples from conditional DDPM or Stable Diffusion can boost classification accuracy in low-data regimes.
• The differences in quality, diversity, and utility of:
  • Vanilla DDPM generations
  • Class-conditional DDPM generations
  • Stable Diffusion (LoRA fine-tuned) generations

---