Cocoa Brain brings major improvements to the meta-verse with new Face-Swap technology

The company’s paper “Smooth-Swap: A Simple Enhancement for Face-Swapping with Smoothness” represents another milestone in Cocoa Brain’s face change research, and will be presented at the upcoming World Vision Conference on Computer, CVPR 2022[2], for the second year in a row. This will include an exclusive oral presentation session reserved for the most outstanding articles among the accepted articles (25.33% of the 8,161 submissions were accepted this year). At last year’s event, only 4% of the accepted articles had time for an oral presentation, where Kakao Brain was nominated for his excellent research article, “HOTR: End-to-End Human-Object Interaction Detection with Transformers”. This year, not only has ‘Smooth-Swap’ significantly reduced the complexity of its architecture, but it also has great market potential, both recognized and awarded by the first conference on computer vision.

An accurate and consistent identity gradient[3] is essential to change a person’s identity without sacrificing high image quality. Trained via monitored contrast loss, ‘Smooth-Swap’ achieves its stable identity gradient by learning incorporation with greater smoothness. These enhancements address the weakness of the previous model by adding handmade components and 3D face modeling, which ultimately complicated its design and resulted in sophisticated hyperparameter adjustment. Instead, “Smooth-Swap” relies on a simple U-Net-based architecture with a built-in identity integrator to deliver industry-leading performance.

The simple architecture and enhanced performance of “Smooth-Swap” have not only made the technology competitive in terms of market potential and wider application, but also enable it to handle face-swap scenarios more challenging, such as face swapping during video playback. “Smooth-Swap” suggests a differentiated identity integration approach and allows the generator to create higher quality images, especially when changing the face shape of a subject. Thanks to Kakao Brain’s “Smooth-Swap”, which enables fast and stable face change, it is expected to develop different types of digital people such as virtual influencers, show hosts and announcers.

“We are proud and excited to unveil the revolutionary ‘Smooth-Swap’ revolutionary face-change technology to the world,” said Kim Il-doo, CEO of Cocoa Brain. “I am convinced that this technology will accelerate innovation in face-sharing and bring us even closer to the incredibly immersive metavers we have always dreamed of and the digital human services of the future.”

About Cocoa Brain

Kakao Brain is a world-leading AI company with unmatched AI technologies and research and development networks. The company was established by Kakao in 2017 to solve some of the world’s biggest “unimaginable issues” with solutions made possible by its life-changing artificial intelligence technologies. Kakao Brain, which is always at the forefront of innovation in the world of technology, has developed many revolutionary AI services and models designed to improve the quality of life for thousands of people, including minDALL-E, KoGPT, CLIP / ALIGN and RQ-Transformer. As a global AI pioneer, Kakao Brain has a responsibility to promote a vibrant technology community and a robust R & D ecosystem as part of its mission to create new technology markets with infinite potential. For more information, visit https://KakaoBrain.com/.

[1] Identity integration is a vector representation of a face image used to compare identities. If the representation vectors (or integration vectors) of two surfaces are sufficiently dense, their identities are considered identical.

[2] CVPR (Conference on Computer Vision and Pattern Recognition), sponsored by the Institute of Electrical and Electronics Engineers (IEEE) and the Computer Vision Foundation (CVF) since 1983, is considered one of the most recognized annual conferences in the computer vision sector, with the European Conference on Computer Vision (ECCV) and the International Conference on Computer Vision (ICCV).

[3] The identity gradient is a training signal that tells the face exchange model which part needs to be tuned to change the person’s identity exactly.

SOURCE Cocoa Brain

rt

Leave a Comment