How AI Restores Blurry Faces: Inside Kuaishou’s Y‑Tech High‑Definition Portrait Project

Background

Image clarity affects many aspects of life, influencing viewing experience, the preservation of precious moments, historical research, child safety, and security investigations. Degraded images result from factors such as capture technique, equipment, imaging systems, and storage/transmission methods, leading to blur, noise, compression, etc.

Project Goal

The aim is to mitigate or eliminate quality loss in portrait images during acquisition and transmission, restoring their true appearance using AI to overcome external limitations, preserve valuable memories, and enhance security.

Problem Analysis and Data Construction

Image degradation can be modeled as a combination of blur, down‑sampling, noise, and compression. The team identified common degradation types and constructed paired high‑quality and low‑quality data by applying random combinations and levels of these degradations.

Noise: Gaussian, shot, impulse

Blur: Gaussian, motion, defocus, resize

Other: JPEG compression, pixelation

By varying proportions, degradation levels, and combinations, realistic paired datasets are generated to simulate real‑world distributions.

Technical Solution

The overall pipeline processes a user‑captured image as follows: face detection and landmark localization, face cropping and alignment, face parsing to obtain a mask, feeding the aligned face and mask into a high‑definition model, and finally compositing the restored face back into the original background.

Model Design

Leveraging the strong structural prior of faces, the face parsing mask is incorporated as a prior in a style‑based generative framework (image→style→image). At each level, a style map derived from the input image and mask modulates the generator’s feature maps, enabling fine‑grained control from coarse to fine details.

Training

The model is optimized using L1 loss, perceptual loss, and adversarial loss, while a data pool introduces diverse degradation types within each batch. Fine‑tuning further reduces dependence on the face parsing model, allowing good results even with poor input quality or inaccurate masks. Background regions also receive denoising, deblurring, and JPEG artifact removal before final compositing.

Conclusion and Outlook

Kuaishou’s Y‑Tech portrait‑HD project restores high‑definition facial details while preserving identity, delivering impressive visual quality and performance now available in the “One Sweet Camera” product. The team will continue to refine algorithms and leverage broader computer‑vision expertise to meet evolving user needs.

References

A Style‑Based Generator Architecture for Generative Adversarial Networks, 2019

Analyzing and Improving the Image Quality of StyleGAN, 2020

stylegan‑encoder, https://github.com/pbaylies/stylegan-encoder, 2019

Image2stylegan: How to embed images into the StyleGAN latent space, 2020

Encoding in Style: a StyleGAN Encoder for Image‑to‑Image Translation, 2021

Additive Angular Margin Loss for Deep Face Recognition, 2018