E-DALL-E: Image Generation and Expansion System

This project combines DALL-E Mini for text-to-image generation with a novel expansion technique (E-DALL-E) that allows creating images with different aspect ratios. It also provides standalone tools for expanding existing images to various aspect ratios without requiring AI generation.

🌟 Overview

The system consists of these integrated components:

1. DALL-E Mini - Generates square (1:1) images from text prompts
2. E-DALL-E - Expands the generated images to different aspect ratios (e.g., 16:9) using VQGAN+CLIP
3. Image Expander (Online) - A tool to download and expand online images to different aspect ratios
4. Image Expander (Local) - A tool to expand local image files to different aspect ratios
5. Results Viewer - A gallery tool to view all expanded images created with the system

📦 Installation

Install the required dependencies:

pip install -r requirements.txt

🚀 Main Features

AI Image Generation and Expansion

Run the main script with a text prompt to generate and expand images:

python edalle.py --prompt "a painting of rolling farmland" --aspect_ratio 1.78 --num_images 4

Command Line Arguments

• --prompt: Text prompt for image generation (default: "a painting of rolling farmland")
• --num_images: Number of images to generate (default: 4)
• --aspect_ratio: Target aspect ratio for expansion (width/height, default: 16/9)
• --iterations: Number of optimization iterations for expansion (default: 100)
• --model_type: DALL-E model size ("mini", "mega", or "mega-full", default: "mega")

Online Image Expansion

Download and expand any online image to your desired aspect ratio:

python download_expand.py --url "https://example.com/image.jpg" --aspect 16:9 --orientation landscape

Example Results

Command Line Arguments

• --url: URL of the image to download (defaults to a sample Van Gogh image)
• --aspect: Target aspect ratio, can be in format "16:9" or decimal "1.78" (default: "16:9")
• --orientation: Orientation of the expansion, "landscape" or "portrait" (default: "landscape")
• --output: Path to save the expanded image (default: "expanded_download.png")

Local Image Expansion

Expand any local image file to your desired aspect ratio:

python local_expand.py input_image.jpg --aspect 21:9 --orientation landscape

Example Results

Command Line Arguments

• First argument: Path to the input image file (required)
• --aspect: Target aspect ratio, can be in format "16:9" or decimal "1.78" (default: "16:9")
• --orientation: Orientation of the expansion, "landscape" or "portrait" (default: "landscape")
• --output: Path to save the expanded image (default: "expanded_" + input filename)

Results Gallery Viewer

Create visual galleries of all expanded images:

python view_results.py

This script will:

1. Scan the current directory for expanded and comparison images
2. Generate two gallery images:
   • comparison_gallery.png: Shows side-by-side comparisons of original and expanded images
   • expanded_gallery.png: Shows all expanded images with their dimensions and aspect ratios

🧠 How It Works

DALL-E + Expansion Process

1. The system first generates square (1:1) images using DALL-E Mini based on your text prompt
2. It then expands one of these images to your desired aspect ratio in two steps:
   • Initially filling in the sides with repeated edge pixels
   • Using VQGAN+CLIP to generate new content on the sides that matches the style and content of the original image and follows the text prompt

Edge Pixel Expansion Technique

The edge pixel expansion process:

1. Creates a canvas with the target dimensions
2. Places the original image in the center
3. Fills the empty regions by repeating and blending edge pixels from the original image
4. Applies a gradient blend to ensure smooth transitions

📋 Supported Aspect Ratios

The system supports any aspect ratio, but here are some common ones:

Aspect Ratio	Common Use Cases
1:1 (1.0)	Standard square images, Instagram
4:3 (1.33)	Traditional TV/monitors
16:9 (1.78)	HD video, modern displays
21:9 (2.33)	Ultrawide monitors, cinematic
2.39:1	Anamorphic widescreen cinema
9:16 (0.56)	Vertical video, mobile devices

🖼️ Output Files

The system generates various output files:

• Original images: original_*.png
• Expanded images: expanded_*.png
• Comparison images: comparison_*.png
• Gallery images: expanded_gallery.png and comparison_gallery.png

🔧 Advanced Usage

Creating Custom Aspect Ratios

You can specify custom aspect ratios in two formats:

• Ratio format: --aspect 4:3
• Decimal format: --aspect 1.33

Combining Tools

You can chain these tools together for creative workflows:

1. Generate images with edalle.py
2. Expand existing images with local_expand.py
3. View all results with view_results.py

📄 License

This project is licensed under the MIT License.

You are free to use, modify, and distribute this software in accordance with the terms of the license.

👏 Acknowledgments

This implementation is based on:

• DALL-E Mini by Boris Dayma and Pedro Cuenca
• VQGAN+CLIP for image expansion
• The original E-DALL-E concept by Robert A. Gonsalves