How Kuaishou’s AI‑Powered Beauty Engine Transforms Real‑Time Video

1. Background of Beauty Technology

Beauty technology removes facial flaws and creates smooth, textured skin. Early tools like Photoshop required expertise, but the rise of live streaming and short videos increased demand for easy, high‑quality beauty effects, leading to AI‑driven one‑click solutions that preserve natural skin texture.

2. Kuaishou Beauty Technology Modules

1. Skin Smoothing

Traditional smoothing relies on edge‑preserving filters such as guided, bilateral, surface, and double‑exponential filters. Different kernel sizes produce varied effects: small kernels smooth fine texture, while large kernels remove blemishes. Combining multiple scales yields both smoothness and natural detail.

Advanced Smoothing

Specialized effects such as “Even Skin” and “Texture Skin” retain facial contours while evening skin tone and preserving texture. “Even Skin” first locates uneven regions and applies targeted filtering; “Texture Skin” uses region‑specific filters to keep fine details while correcting defects.

2. Facial Blemish Removal

Blemishes such as acne, moles, and spots require detection and repair beyond smoothing. Kuaishou employs a CNN with a multitask segmentation head to locate and classify blemishes, a FeatureAnalyzer for attribute extraction, and a PriorityNet that generates attention maps guiding selective restoration. The repair stage uses an inpainting‑style network to fill removed regions with realistic skin texture.

On an iPhone 6 the end‑to‑end pipeline runs in about 60 ms, outperforming competing solutions and remaining robust under low‑light or back‑light conditions.

3. Clarity Enhancement

The clarity module boosts visual detail while mitigating typical sharpening artifacts such as white or black edges, delivering sharper yet natural‑looking images.

4. Skin Tone Adjustment

Supports warm, cool, white, and black tone targets. The algorithm uses LUT‑based color grading combined with edge diffusion to ensure smooth transitions between facial skin and background.

5. Local Facial Fine‑Tuning

Local adjustments enhance eyes (sharpening or overlay) and teeth (whitening). Teeth whitening leverages a Generative Adversarial Network to produce realistic, aligned results.

6. AI Beauty

Kuaishou’s AI beauty solution, GorgeousGAN, follows a UNet‑style architecture with three stages: feature extraction (down‑sampling convolutions), feature processing (dilated convolutions forming a feature pyramid and attention via the proprietary STAModule), and image generation (noise injection for realistic skin texture). Training losses include global adversarial loss and LPIPS, plus a dedicated local discriminator for facial features. A separate classifier adjusts facial depth based on lighting conditions, ensuring consistent 3‑D effect across diverse scenes.

3. Conclusion and Outlook

The Gorgeous platform integrates traditional image‑processing pipelines with cutting‑edge AI models to provide a wide range of beauty effects, from basic smoothing to sophisticated AI‑driven enhancements. Future work will focus on expanding the diversity of beauty styles and further advancing the underlying technologies to deliver ever more compelling user experiences.