How Kuaishou Achieves Realistic Body Beautification with AI‑Driven Pose Detection and Image Warping

Background

With the rise of short‑video platforms, portrait beautification has become a standard feature in many apps. Leveraging Kuaishou’s self‑developed human pose detection and key‑point localization technologies, the Y‑tech team has released seven body‑beautification effects—"Long Leg", "Small Head", "Swan Neck", "Slim Waist", "Full Chest", "Slim Shoulder", and "Slim Thigh"—available in Kuaishou, Kuaishou Short Video, and other apps.

Body‑Beautification Technology Overview

Human Pose Detection & Key‑Point Localization

Accurate and stable body‑beautification starts with reliable human pose detection and key‑point localization. For effects such as "Swan Neck" or "Slim Shoulder", precise shoulder points are required; for "Slim Waist", accurate waist points are essential. Kuaishou’s proprietary key‑point model provides the necessary precision.

Deformation Algorithms Used in Body‑Beautification

The pipeline employs three main image‑warping algorithms: image stretching, triangulation, and liquify. Image stretching expands a region according to a predefined rule, suitable for large‑scale changes. Triangulation divides the target area into triangles and applies affine transformations for fine‑grained control. Liquify moves pixels around a control point to achieve smooth, rounded deformations.

Image Stretching Algorithm

For the "Long Leg" effect, the region below the hip line is stretched toward the foot direction after key‑point detection. The stretching factor λ(y,c) provides a smooth transition, preventing abrupt visual artifacts.

Formula (simplified): P2 = P1 + λ * (target_line_midpoint - P1)

Triangulation Algorithm

Triangulation creates a mesh of triangles over the target region. Each triangle undergoes an affine transformation derived from the before‑and‑after vertex coordinates. This enables precise control of deformation for areas such as the thigh.

Liquify Algorithm

Liquify defines a circular influence area centered on a control point. Pixels inside the circle are pushed toward a direction vector, with intensity decreasing toward the boundary. This method is used for effects such as "Slim Waist" and "Swan Neck".

Divide the image into a fine mesh.

Detect key points (e.g., waist points P1 and P2).

Set the waist line as the liquify direction and define a radius.

Move mesh points inside the radius according to the liquify formula.

Render the final "Slim Waist" effect.

Advantages of Kuaishou’s Body‑Beautification Technology

Background Distortion Correction

When deforming a human figure, the background often warps, leading to unnatural results (e.g., “the wall is crooked”). Kuaishou’s self‑developed background‑distortion correction algorithm keeps the slope of background lines unchanged, preserving visual consistency.

For the "Small Head" effect, the correction removes noticeable background warping, as shown in the comparison.

Reduced Background Influence

By leveraging precise key‑point data and adaptive algorithm parameters, Kuaishou minimizes the impact on the background. Effects such as "Swan Neck" and "Slim Thigh" automatically adjust liquify radius and intensity, while a dynamic strength‑reduction mechanism further limits unintended deformation.

High Visual Stability

In video scenarios, rapid body movement can cause background jitter. Kuaishou incorporates motion‑aware constraints and smooth transition functions to keep the background stable, resulting in smoother playback compared with competing solutions.

Overall, the body‑beautification pipeline combines accurate pose detection, robust deformation algorithms, and background‑preserving techniques to deliver natural, stable, and user‑friendly visual effects.

References

Andreas Gustafsson, Interactive Image Warping.

Key‑point detection in Kuaishou: https://mp.weixin.qq.com/s/ab789Vd74DW1sQT7duxmSg

Transformations: http://www.cs.tau.ac.il/~dcor/Graphics/cg-slides/trans3d.pdf

R. Gioi et al., “LSD: A fast line segment detector…”, IEEE TPAMI, 2010.

A. Vakhitov & V. Lempitsky, “Learnable line segment descriptor…”, IEEE Access, 2019.

C.-H. Chang et al., “Shape‑preserving half‑projective warps…”, CVPR, 2014.