This AI makes blurry faces look 60 times sharper! Introduction to PULSE: photo upsampling

Introduction

In recent advancements in artificial intelligence, a new algorithm has emerged that can transform a blurry image into a high-resolution version, turning super low-resolution images (like 16x16 pixels) into stunning 1080p HD human faces. It might sound unbelievable, but you can try it yourself in less than a minute. Before diving into your own testing, let’s explore how this remarkable technology works.

Understanding Photo Upsampling and Image Super Resolution

To start, we must grasp the concepts of photo upsampling and image super resolution. The goal of this technology is to create a high-resolution image from a significantly lower-resolution input, specifically focusing on faces in this context. The low-resolution input can be as minimal as a 16x16 pixel image, which is often super blurry. The challenge is reconstructing this input into a crisp, clear 1080p image.

Traditionally, these techniques employ supervised learning to train networks, measuring the average distance between the generated high-definition image and a high-resolution ground truth. However, this approach often neglects essential details like textures, leading to blurry spots where the algorithm is uncertain about how to proceed due to differences in resolution.

Enter PULSE

This is where PULSE (Photo Upsampling via Latent Space Exploration) comes into play. The PULSE algorithm aims to generate realistic images from a set of plausible solutions. Instead of relying solely on the low-resolution input to guess what the high-resolution image should look like, it utilizes an actual high-resolution image that can realistically downscale to the original low-resolution input.

PULSE employs a self-supervised technique that navigates through the manifold of high-resolution natural images, searching for images that can be downscaled to match the original low-resolution image. This is beneficial because multiple high-resolution images can correspond to the same low-resolution image, allowing the algorithm to produce more accurate results without the blurry uncertainty that earlier methods exhibited.

The high-resolution images generated are produced using a Generative Adversarial Network (GAN), which has been pre-trained in an unsupervised manner to create multiple realistic and sharp face images.

Initially skeptical, I decided to put this algorithm to the test myself, as the code is publicly available. The results were astounding! Even from a super-blurry image where my identity was hard to discern, the algorithm generated a remarkably close approximation of my face. You can easily try it with your own image in just a minute. The program is easy to use, and your privacy is guaranteed as your images won’t be saved.

PULSE utilizes the CelebHQ dataset, a large-scale face attribute dataset comprising over 200,000 celebrity images. Consequently, the faces produced often appear older and distinct compared to non-celebrity faces. This raises interesting questions about the algorithm's efficacy with different demographics.

This innovative work holds great potential in various fields, including medicine, astronomy, and satellite imagery, where precise, high-resolution images are traditionally challenging to obtain due to costs or hardware limitations.

To learn more about PULSE and its intricacies, I wholeheartedly recommend reading the original paper and experimenting with the code, both of which can be found linked in the description below.

Keywords

• PULSE
• Photo Upsampling
• AI
• Image Super Resolution
• Generative Adversarial Networks (GAN)
• CelebHQ dataset
• Self-Supervised Learning
• High Resolution
• Machine Learning

FAQ

Q1: What is PULSE?
A1: PULSE is an AI algorithm designed to upscale low-resolution images, transforming them into high-resolution versions by relying on realistic downscaled images.

Q2: How does PULSE work?
A2: PULSE utilizes a self-supervised learning technique that searches through a set of high-resolution images to find those that can be downscaled to match a given low-resolution image.

Q3: Can I try PULSE for myself?
A3: Yes, the code for the PULSE algorithm is available online, and you can input your own images to see the results—no setup required.

Q4: What type of images can PULSE enhance?
A4: PULSE is particularly geared towards enhancing human faces in images, but its applications can extend to other fields such as medicine and astronomy.

Q5: What dataset does PULSE use?
A5: PULSE utilizes the CelebHQ dataset, which contains over 200,000 celebrity images, to train its model for generating realistic and sharp faces.