Kakao Brain announces new face-swapping technology ‘Smooth-Swap’

Kakao Brain has reportedly introduced Smooth-Swap, its latest face-swapping technology, developed to extend the abilities of smooth face-swapping through improved identity embedding.

The innovative model consists of simplified, ground-breaking architecture after exchanging complicated external modules of current systems with a consolidated smooth identity embedder that facilitates more stable, faster face-swapping.

Kakao Brain’s paper ‘Smooth-Swap: A Simple Enhancement for Face-Swapping with Smoothness’ will be showcased at CVPR 2022, the forthcoming global computer vision conference, for two years in a row. This development represents another crucial achievement in the company’s face-swapping research.

The event will also include an exclusive session of oral presentations reserved for the most exceptional papers amongst accepted articles.

In 2021’s event, a mere 4% of accepted papers were allotted time to conduct an oral presentation in which Kakao Brain was selected for its outstanding research paper, HOTR: End-to-End Human-Object Interaction Detection with Transformers.

This year, the Smooth-Swap technology has not only managed to substantially reduce the complexity of its architecture, but also has the immense ability for commercialization, both of which have been identified and awarded at the conference.

A consistent and precise identity gradient is important to smoothly change a person’s identity without compromising high image quality. Smooth-Swap trains through supervised contrastive loss and obtains its identity gradient by smoothly learning to embed.

These enhancements acknowledge the previous model’s weakness of including 3D face modelling and handcrafted components which eventually complicated its design and implied sophisticated tuning of hyperparameters.

Smooth-Swap relies on a simple architecture based on U-Net with a consolidated smooth identity embedder to provide top performance.

The improved performance and simple architecture of Smooth Swap have not only made it commercially competitive but also enabled it to face more complex face-swapping scenarios like swapping a face during video playback.

Smooth-Swap recommends a differentiated approach to embedding identity and authorizes the generator to create high-quality images, particularly while changing the face shape of a subject.

Source credits: https://www.prnewswire.com/news-releases/kakao-brain-brings-big-improvements-to-the-metaverse-with-new-face-swap-technology-301541582.html