Diffusion Model

A powerful conditional diffusion model for image generation from text descriptions.

---

Overview

This project implements a conditional diffusion model that generates images from text descriptions. The model learns to gradually denoise random Gaussian noise into coherent images, guided by text prompts.

Key Features

• Text-to-Image Generation: Convert textual descriptions into high-quality images
• Web Interface: User-friendly interface for image generation
• Conditional Generation: Fine-grained control over generated content
• Interactive Process Visualization: Watch the denoising process in real-time
• Multi-Class Support: Generate images across various categories

Architecture

The model consists of several key components:

├── models/
│   ├── modules.py        # Neural network building blocks
│   ├── text_encoder.py   # Text embedding module
│   ├── time_encoder.py   # Timestep encoding module
│   └── unet.py           # Conditional UNet architecture
│
├── output_ImageNet/      # Generated outputs and visualizations
│
├── diffusion.py          # Core diffusion model implementation
├── plot_func.py          # Visualization utilities
├── preprocess.py         # Data loading and preprocessing
└── web.py                # Web interface for image generation

Web Interface

The project includes a sleek web interface for easy interaction with the model:

Features:

• Text prompt input
• Multiple image generation
• Process visualization
• Real-time generation progress

Model Components

1. Diffusion Process

• Forward diffusion adds noise gradually
• Reverse diffusion learns to remove noise
• Conditional generation guided by text embeddings

2. Architecture Details

• UNet backbone with skip connections
• Text conditioning through cross-attention
• Time embedding using sinusoidal positions
• Batch normalization for stable training

3. Dataset

• Tiny ImageNet, over 1Million images and 200 classes

3. Training Process

• Dataset: ImageNet subset
• Text embeddings: SentenceTransformer
• Loss: MSE between predicted and actual noise
• Optimizer: AdamW with gradient scaling

Sample Generations

Generate images from text descriptions like:

• "red apple"
• "golden retriever"
• "sunset over mountains"

License

This project is licensed under the MIT License - see the LICENSE file for details.

---