Magic The Gathering - Cards Image Segmentation

This project provides tools for generating Magic The Gathering Cards image segmentation machine learning model.

• Pre-generated dataset (step 1) in this
• Trained model (step 2) in this link
• Demo (step 3) in this link

Glossary

• Magic The Gatering Good card game for the cool people!
• Machine Learning A field of artificial intelligence
• Python Programing Language
• Blender 3D modeling and rendering Open Source software
• BlenderProc2 A procedural Blender pipeline (tool) for photorealistic rendering
• HDRI High Dynamic Range Image, for creating realistic environments
• Skryfall MTG card database
• Polyhaven Website for 3d assets
• API An interface for talking with websites/apps
• ONNX Open Neural Network Exchange
• GAN Generative Adversarial Network (IA image generators for example)

The Plan

• Create MTG Cards Syntetic Dataset Generator based on 3d redering, and generate the first version of this dataset.
• Create the machine learning training scrips and train the model
• Transform to ONNX
• Crate a web example for runing the model in web

Hardware/Software Requirements

• 16 gb of ram or more (recomended)
• 10 GB of hard disk space (or more)
• A cuda 12 capable GPU (Nvidia) (others can be used, not tested)
• A working pytorch environment with GPU acceleration (google it)
• Ubuntu 24.04 is recomended

Step 1 - MTG Cards Syntetic Dataset Generator (dataset_generator/)

NOTE: You cand download pregenerated dataset from this link.

1. Reference Dataset Downloader (01_download_references.py)

Downloads MTG card reference images from Scryfall API.

2. Reference Dataset Downloader (02_dowload_hdrs.py)

Downloads realistic environments and lightings (HDRI) from Polyhaven API.

3. Synthetic Image Generator (03_generate_synthetic_dataset.py)

BlenderProc2-based synthetic image generator that creates realistic MTG card images with segmentation masks.

Features:

• Photorealistic rendering using HDRI environments
• Precise card geometry with rounded corners
• Random transformations (0-15° rotation on all axes)
• Variable positioning (40-70% image coverage)
• Automatic segmentation masks generation

Examples:

Quick Start:

# Create and activate python virtual enviroment
python3 -m venv venv
source venv/bin/activate

# Upgrade pip package manager 
python -m pip install --upgrade pip

# Install dependencies
python -m pip install -r requirements.txt
blenderproc pip install tqdm>=4.66.0

# Download reference images
cd dataset_generator
python 01_download_references.py

# Download enviroments
python 02_dowload_hdrs.py

# Generate synthetic images
cd ..
blenderproc run dataset_generator/03_generate_synthetic_dataset.py

# For large datasets, use synthetic_dataset_anti_leak.py (restar the 03_generate_synthetic_dataset script every X minutes)
python synthetic_dataset_anti_leak.py -t 5

# blenderproc recomend not redering too many imgs in a single bender instance. I read this after the implementation...
# So... i make a dirty hack (it works!) 

Note: In Windows, if blenderproc is not in your path, you can try (examples):

python -m blenderproc pip install tqdm>=4.66.0
python -m blenderproc run dataset_generator/03_generate_synthetic_dataset.py

Step 2 - Train (train/)

NOTE: You cand download the trained model form this link.

Please read this documentation link.

Step 3 - Demo (demo/)

NOTE: You cand access this demo form this link.

Quick Start:

pip install fastapi uvicorn
python demo.py
INFO:     Started server process [5926]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:5000 (Press CTRL+C to quit)

Go to http://127.0.0.1:5000/ in you browser.

Note: You need the exported model in train/exported_models/card_segmentation.onnx (see step 2)

Developer Notes

This tools where created using:

VSCode
Cline - AI Dev Assistant / VSCode extension
Copilot - IA Dev Asistant (provider) / VSCode extension
Claude 4 - IA Model
Gemini 2.5 pro - IA Model
Python / Python Debugger / Pylance - VSCode extensions

See Synthetic Image Generator documentation.
Original repo link.

Sorry for my english, im still learning.