Synthetic Image Generation with GAN (DCGAN)

Generate synthetic (“fake”) images using a Generative Adversarial Network (GAN).
This project trains a DCGAN to learn the distribution of a dataset and produce new, realistic-looking samples from random noise.

Note: These images are synthetic, created by the model — not copied from the dataset.

---

✨ Project Highlights

• ✅ DCGAN training loop (Generator vs Discriminator)
• ✅ Stabilization tricks (label smoothing, noise injection optional)
• ✅ Saves generated image grids during training
• ✅ Checkpoint saving + inference script

---

🧠 How GAN Works (Quick)

• Generator (G): takes random noise z → produces a fake image
• Discriminator (D): receives an image → predicts real vs fake
• Training is adversarial: G tries to fool D, D tries to catch G

---

📦 Dataset

This repo is set up for:

• Fashion-MNIST (default, 28×28 grayscale)
  You can switch to CIFAR-10 or custom datasets with minor changes.

---

🏗️ Model Architecture (DCGAN)

Generator

• Dense → reshape → Conv2DTranspose blocks → output image (tanh)

Discriminator

• Conv2D blocks → flatten → sigmoid output (real/fake)

---

🚀 Getting Started

1) Clone the repo

git clone <YOUR_REPO_URL>
cd <YOUR_REPO_NAME>