MalumaDev/VAEAEDOF

VAEEDOF - High-Resolution Multi-Focus Image Fusion

Model Description

VAEEDOF is a deep learning model designed to address the Depth-of-Field (DOF) constraint in photography using Multi-Focus Image Fusion (MFIF). Built upon a distilled Variational Autoencoder (VAE) architecture, this model fuses up to 7 images with different focus points into a single, high-resolution, all-in-focus image.

It is trained to produce artifact-free and photorealistic fused outputs and demonstrates strong generalization across both synthetic and real-world datasets.

📦 Model Weights

This repository provides:

• ✅ Pretrained VAEEDOF weights used in our experiments

• 🔁 Comparison model weights for evaluating against other state-of-the-art methods (baselines)

🧪 Training Data

The model is trained on the MattingMFIF dataset – a new, high-quality 4K synthetic dataset built using matting techniques applied to real-world photographs to simulate realistic depth-of-field blur and focus patterns.

🔗 Resources

GitHub Repository (Code, training & inference scripts): 👉 https://github.com/MalumaDev/VAEEDOF

📄 Citation

@article{piano2025addressing,
  title={Addressing the Depth-of-Field Constraint: A New Paradigm for High Resolution Multi-Focus Image Fusion},
  author={Piano, Luca and Huanwen, Peng and Bilcu, Radu Ciprian},
  journal={arXiv preprint arXiv:2510.19581},
  year={2025}
}