🧠 Human Face Image Generation & Completion

Applied ML Final Project
Contributors: Ananya Sinha, Ankita, Divyanshi Kumari, Rohit Roy

---

🚀 Overview

This project uses Deep Convolutional GANs (DCGANs) to:

• Generate realistic human face images.
• Complete masked regions (e.g., mouth/nose) using contextual and perceptual losses.

---

📂 Dataset

CelebA Dataset

• 202,599 RGB images (80×80 px)
• Filtered 91k+ straight-facing images
• Applied 30×30 pixel mask near mouth region

---

🧠 Model Architecture

Generator:

• Input: 100-dim noise
• Layers: Dense → Upsampling → Convs → tanh

Discriminator:

• Convs → MaxPooling → Dense → sigmoid

---

🛠 Methodology

1. Mask part of the face.
2. Optimize latent vector to minimize combined loss:
   • Contextual (pixel match)
   • Perceptual (realism via discriminator)
3. Merge generated and original pixels.

---

📉 Results

• Best results at 9 epochs.
• Small masks (30×30): realistic outputs.
• Large masks (40×40): loss of detail.
• More iterations = better tone but worse structure.

---

⚠️ Challenges

• Training instability (common in GANs)
• Sensitive to architecture tweaks
• tanh activation worked best

---

🔮 Future Work

• Use advanced models like StyleGAN
• Irregular/flexible masking
• Better preprocessing & transfer learning

---

📎 Links

• CelebA Dataset