3Dify: 2D to 3D Stereoscopic Image Converter

3Dify is a powerful Python application that converts standard 2D images into immersive 3D stereoscopic views. By leveraging state-of-the-art local AI depth estimation models and custom stereo-generation algorithms, it allows you to dynamically generate and customize 3D effects from any flat image.

🌟 Key Features

1. State-of-the-Art Depth Estimation

The project uses the depth-anything/Depth-Anything-V2-Small-hf model via the Hugging Face transformers pipeline. This provides highly accurate depth maps where foreground objects and background landscapes are distinctly separated, ensuring a clean 3D effect. The pipeline automatically utilizes GPU acceleration (cuda or mps) if available, falling back to CPU if not.

2. Advanced Stereo Generation Algorithm

Rather than simple pixel shifting, the core engine in stereo_generation.py uses a custom inverse mapping technique with occlusion handling:

• Disocclusion Rejection: It detects "stretched" pixels on steep depth slopes (where depth decreases sharply) and rejects them.
• Background Hole Filling: Gaps created by foreground objects shifting are intelligently filled. The algorithm scans horizontally to sample pure background pixels (offsetting to avoid bleeding from noisy foreground outlines), preventing ugly stretching artifacts.
• Bilateral Smoothing: Applies OpenCV bilateral filtering (cv2.bilateralFilter) to the depth map to smooth noise while preserving sharp object edges.

3. Fully Customizable Parameters

Through the interactive UI, you can tweak the 3D generation to perfection:

• Stereo Strength: Controls the intensity of the 3D effect (i.e., how far apart the left/right views are shifted based on the depth map).
• Maximum Disparity: The absolute maximum horizontal pixel shift applied to the closest foreground objects.
• Depth Smoothing: Size of the bilateral filter to reduce noise on depth boundaries.
• Inpainting Radius: Controls the heuristic for hole-filling pixel-shifted occluded areas.

4. Multiple Output Formats

Whether you are using a VR headset, a 3D TV, or just red/cyan glasses, this tool supports multiple formats:

• Anaglyph (Red/Cyan)
• Side-by-side (SBS)
• Half Side-by-side
• Over-under (Top/Bottom)
• Separate (Left/Right views)

5. Interactive Web Interface

Built with Gradio, the UI provides immediate visual feedback. You can view the final 3D output alongside intermediate steps like the raw depth map and the isolated Left/Right eye views.

---

🚀 Installation & Setup

This project requires Python 3.10+.

Step 1: Clone the Repository

git clone https://github.com/arcaneleague18/3Dify.git
cd 3d

Step 2: Install Dependencies

It is highly recommended to use a virtual environment. The project is pre-configured for uv (a fast Python package installer and resolver), but standard pip works perfectly as well.

Using uv (Recommended):

uv sync

Using standard pip:

python -m venv .venv
# Activate virtual env:
# Windows: .venv\Scripts\activate
# Mac/Linux: source .venv/bin/activate

pip install -r src/requirements.txt

---

🎮 Usage

To launch the Gradio web interface, simply run:

python src/main.py

(Depending on your entry point, you can also run python src/ui.py directly).

1. Open the provided local URL (e.g., http://127.0.0.1:7860) in your web browser.
2. Upload a standard 2D image.
3. Adjust the stereo parameters (Strength, Disparity, Smoothing).
4. Select your desired output format from the dropdown.
5. Click "Generate 3D Image" and view your results!

---

📁 Repository Structure

• src/ui.py: Gradio interface setup and event handling.
• src/depth_estimation.py: Inference code for generating depth maps using transformers.
• src/stereo_generation.py: Core algorithm for inverse pixel mapping and occlusion hole-filling.
• src/image_utils.py: Utilities for formatting the left and right eyes into Anaglyph, SBS, Over/Under, etc.
• src/inpainting.py: Additional functions for gap filling (if applicable).
• src/config.py: Global constants, model names, and default parameters.
• src/main.py: Entry point for the application.