Here are the most important scientific articles. It's a hand-curated selection of the latest AI and data science advancements, presented chronologically with a link to a more detailed article.

Photorealism Enhancement

Researchers can use this AI in real-time to make every video game frame look more natural. Enhancing Photorealism Enhancement is the name of a paper that researchers from Intel Labs just put out. And if you think this is "just another GAN" that takes a picture from a video game and changes it to look like the real world, let me change your mind. They worked for two years on this model to make it very strong. It can be used live in a video game to make every frame look more natural.

The researchers show a way to make computer-made images look more natural. A convolutional network uses intermediate representations made by traditional rendering pipelines to improve the pictures. The network has a new adversarial goal, which gives it intense supervision on many levels. The researchers look at how scene layouts are distributed in commonly used datasets and find essential differences. They think that this is one of the reasons why the results of many previous methods show powerful artefacts.

The researchers have developed a new way to sample image patches while training to solve this problem. They also make several architectural changes to the deep network modules used to improve photorealism. In controlled experiments, the researchers confirm the benefits of their work and report significant improvements in stability and realism compared to recent image-to-image translation methods and other baselines.

Paper: Enhancing Photorealism Enhancement

Click here for the code

Deepfake Detector

In this work, researchers propose DefakeHop, a method for finding Deepfakes that is both light and fast. Deep neural networks are the foundation of the most advanced ways to find Deepfakes. DefakeHop automatically pulls features from different parts of face images by using the successive subspace learning (SSL) method. 

The c/w Saab transform is to pull out the features. Then their feature distillation module uses spatial dimension reduction and soft classification for each channel to make a shorter description of the face. Moreover, DefakeHop has a small model size, a fast training process, and a high detection AUC, and it needs fewer training samples than other methods. Researchers did many tests to show that it works well at detecting things.

Paper: DefakeHop: A Light-Weight High-Performance Deepfake Detector

Photorealistic Image Translation

Existing image-to-image translation (I2IT) methods can only work with low-resolution images or take along to decide. They have to do a lot of work to combine high-resolution feature maps.

In this paper, the researchers focus on how to speed up high-resolution photorealistic I2IT tasks based on closed-form Laplacian pyramid decomposition and reconstruction. In particular, the researchers show that attribute changes, such as changing the lighting and colours, have more to do with the low-frequency component. At the same time, researchers can refine the details of the content adaptively on the high-frequency component.

So, the researchers suggest a Laplacian Pyramid Translation Network (LPTN) do both of these things simultaneously. They design a lightweight network to translate the low-frequency component with less accuracy and a progressive masking strategy to refine the high-frequency components quickly. Their model eliminates the most complicated calculations needed to process high-resolution feature maps and keeps the exact image details. Extensive testing on various tasks shows that the proposed method can translate 4K images in real-time using just one standard GPU, with performance comparable to that of existing methods.

Paper: High-Resolution Photorealistic Image Translation in Real-Time: A Laplacian Pyramid Translation Network

Click here for the code.