Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration

Guy Ohayon ,  Tomer Michaeli ,  Michael Elad

[Paper] |  [Project Page] |  [Code]

Gradio demo for the blind face image restoration version of Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration. You may use this demo to enhance the quality of any image which contains faces.

PMRF is a novel photo-realistic image restoration algorithm. It (provably) approximates the optimal estimator that minimizes the Mean Squared Error (MSE) under a perfect perceptual quality constraint. Our model in this demo is specifically tailored for blind face image restoration. Please refer to our project's page for more details: https://pmrf-ml.github.io/.

Notes :

  1. Our original model is designed to restore low-quality face images, where the image is square, there is only one face in the image, and the face is centered and aligned. In this demo, however, we incorporate mechanisms that allow restoring the quality of any image that contains any number of faces. Thus, the resulting quality of such general images is not guaranteed.
  2. If your image is not an aligned and square face image, make sure that the checkbox "The input is an aligned and square face image" in not marked.
  3. Too large images may result in out-of-memory error.
1 4
1 200
0 10000
Examples
Input Randomize seed The input is an aligned and square face image Scale factor (applicable to non-aligned face images) Number of inference steps (a larger number should lead to better image quality) Seed

If you find our work useful, please ⭐ our GitHub repository. Thanks!

📝 Citation

@article{ohayon2024pmrf,
  author    = {Guy Ohayon and Tomer Michaeli and Michael Elad},
  title     = {Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration},
  journal   = {arXiv preprint arXiv:2410.00418},
  year      = {2024},
  url       = {https://arxiv.org/abs/2410.00418}
}

📋 License

This project is released under the MIT license.

📧 Contact

If you have any questions, please feel free to contact me at guyoep@gmail.com.