This repository contains a comprehensive collection of implementations and research reports for the Deep Generative Models course at the University of Tehran. The projects explore the theoretical foundations and practical applications of modern generative architectures, ranging from Variational Autoencoders (VAEs) to Diffusion Models.
Each project includes a dedicated report, mathematical derivations, and a self-contained Jupyter Notebook implementation.
| Project | Description | Key Topics |
|---|---|---|
| 1. VAE & Disentanglement | Analysis of latent variable independence using -VAE on the dSprites dataset. Includes custom implementations of the MIG (Mutual Information Gap) metric to quantify disentanglement. | VAE Beta-VAE Disentanglement MIG Score |
| 2. Normalizing Flows & GANs | Implementation of Masked Autoregressive Flows (MAF) for industrial anomaly detection (MVTec AD) and CycleGAN for unpaired style transfer (Summer Winter). | Normalizing Flows MADE CycleGAN Anomaly Detection |
| 3. Energy & Score-Based Models | Exploration of implicit likelihood models. Features an Energy-Based Model (EBM) with a replay buffer and a Noise Conditional Score Network (NCSN) for conditional digit generation. | EBM Langevin Dynamics Score Matching NCSN |
| 4. Diffusion & Flow Matching | Implementation of DDPM & DDIM from scratch for FashionMNIST generation, parameter-efficient fine-tuning of Stable Diffusion via DreamBooth & LoRA, and Conditional Flow Matching for synthesizing financial time series (SPY ETF). | Diffusion DDPM/DDIM DreamBooth Flow Matching |
Note: Please refer to the individual project folders for detailed READMEs.