MILO

This is the official repository for:

MILO: A Lightweight Perceptual Quality Metric for Image and Latent-Space Optimization
Ugur Çoğalan, Mojtaba Bemana, Karol Myszkowski, Hans-Peter Seidel, Colin Groth
ACM Transactions on Graphics (TOG), 2025

🔗 Official Project Page: milo.mpi-inf.mpg.de

---

Features

• Full-Reference Image Quality Assessment (FR-IQA)
  • Predicts perceptual quality scores aligned with human judgments.
  • Outputs both a global MOS score and a spatial visibility map for distortion localization.
  • Achieves state-of-the-art accuracy with a lightweight, multiscale CNN.
  • Trained using pseudo-MOS labels, eliminating the need for large-scale human annotation.
• Latent-Space Optimization and Perceptual Loss
  • Operates directly on VAE-encoded latent representations to guide diffusion-based pipelines.
  • Provides perceptually aligned optimization without repeated decoding into image space.
  • Enables curriculum learning: first restores less critical regions, then focuses on perceptually salient areas.
  • Improves performance in tasks such as denoising, super-resolution, and face restoration.

Usage

To run the MILO metric, first clone this repository. Then simply execute:

python MILO_runner.py --ref images/ref.png --dist images/dist.png

• --ref : Path to the reference image
• --dist: Path to the distorted/test image

The output score indicates the perceptual difference:

• Score of 0 → perceptually identical images (no visible distortions)
• Score of 1 → highest possible disruption; extreme perceptible distortions in the test image (compared to the reference)

To run the MILO latent metric:

python MILO_runner_latent.py --ref latents/ref.npy --dist latents/dist.npy

The .npy files contain the latent space representation of the reference and distorted images.

Installation

Requirements

• Python 3.8+
• PyTorch (with CUDA enabled)
• torchvision

Installation Command

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

Citation

If you use MILO in your research, please cite:

@article{Cogalan2025MILO,
  author    = {Ugur {\c{C}}o{\u{g}}alan and Mojtaba Bemana and Karol Myszkowski and Hans-Peter Seidel and Colin Groth},
  title     = {MILO: A Lightweight Perceptual Quality Metric for Image and Latent-Space Optimization},
  journal   = {ACM Transactions on Graphics (TOG)},
  year      = {2025},
  volume    = {44},
  number    = {6},
  publisher = {ACM},
  doi       = {10.1145/3763340}
}

You may also refer to our earlier paper: https://github.com/ugurcogalan06/Enhanced-IQM/

Acknowledgments

This project was produced by research at the Max Planck Institute for Informatics, Germany.