Technical Overview of Kuaishou Y‑Tech Body‑Shaping Effects and Underlying Algorithms

With the rise of short‑video platforms, portrait beautification has become a standard feature; Kuaishou Y‑Tech provides seven body‑shaping effects (e.g., long‑leg, small‑head, swan‑neck, slim‑waist, bust‑enhance, slim‑shoulder, slim‑thigh) that are powered by proprietary human detection and keypoint localization technologies.

2. Body‑Shaping Technology Introduction

2.1 Human Detection & Keypoint Localization

Accurate and stable body‑shaping requires precise detection of human bodies and keypoints; for example, the "swan‑neck" effect needs reliable shoulder points, while "slim‑waist" relies on waist points. Kuaishou's self‑developed keypoint model supplies these coordinates, as illustrated in Figure 2.

2.2 Deformation Algorithms Used in Body‑Shaping

2.2.1 Image Stretching

The stretching algorithm scales a selected region according to a rule; for the "long‑leg" effect, the area below the hip line is stretched toward the foot direction, using a smooth gradient function λ(y,c) to control deformation intensity.

2.2.2 Triangulation (Affine Mesh) Algorithm

Triangulation divides the target region into triangles and applies affine transforms to each mesh, enabling fine‑grained control of deformation. The "slim‑thigh" effect demonstrates how a well‑designed mesh yields precise shape changes.

2.2.3 Liquify Algorithm

Liquify uses a control point as the center of a circular influence region; pixels inside are pushed toward a direction with intensity governed by a curve g. The "slim‑waist" effect applies this by centering on waist points, defining a radius, and moving interior mesh points accordingly.

2.3 Advantages of Kuaishou's Body‑Shaping Technology

2.3.1 Background Distortion Correction

To avoid unnatural background warping (e.g., crooked walls), Kuaishou introduces a background‑distortion correction algorithm that keeps line slopes unchanged and preserves triangle shapes, dramatically reducing visual artifacts as shown in Figures 12‑13.

2.3.2 Limited Background Influence

By adapting liquify radius and curve based on body part characteristics and employing a dynamic strength reduction mechanism, Kuaishou minimizes background impact, achieving more realistic results compared with competitors (Figures 14‑15).

2.3.3 High Visual Stability

Through motion‑aware constraints and smooth gradient functions, the system maintains stable frames even during vigorous movements, preventing background jitter that is common in other solutions (Figures 16‑18).

3. Conclusion

Body‑shaping effects are essential for portrait beautification; Kuaishou's Y‑Tech combines precise human keypoint detection with robust deformation algorithms and background correction to deliver natural, stable results, and will continue to evolve with user aesthetic demands.

References

[1] Andreas Gustafsson, Interactive Image Warping

[2] https://mp.weixin.qq.com/s/ab789Vd74DW1sQT7duxmSg (Keypoint detection in Kuaishou)

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

[4] Rafael Gioi et al., LSD: A fast line segment detector, IEEE TPAMI, 2010

[5] A. Vakhitov and V. Lempitsky, Learnable line segment descriptor for visual SLAM, IEEE Access, 2019

[6] C.-H. Chang et al., Shape‑preserving half‑projective warps for image stitching, CVPR, 2014