How Kuaishou Y‑Tech Achieves Real‑Time High‑Fidelity Wrinkle Removal with AI

Wrinkles are the earliest visible signs of skin aging, typically appearing after age 25 in a predictable order across facial regions, and their removal has long been a challenging problem in beauty technology.

Traditional patch‑matching based de‑wrinkling methods often produce blurry textures and artifacts. Kuaishou Y‑Tech integrates image inpainting [2,9] with semantic image editing [3] to achieve accurate, high‑realism wrinkle removal, as demonstrated by the neck‑wrinkle results in Figure 1.

Technical Solution

The overall training pipeline (Figure 2) and inference pipeline (Figure 3) illustrate the end‑to‑end process of the smart de‑wrinkling system.

Overall Framework

During training, facial images are first segmented using Kuaishou’s proprietary wrinkle segmentation to obtain masks for different wrinkle categories. Random wrinkle‑shaped masks are generated in non‑wrinkled skin regions, the masked pixels are removed, and the resulting images are fed into a self‑supervised network to learn the missing information.

During inference, the same segmentation provides wrinkle masks, which are erased and then processed jointly by an image‑inpainting network and a semantic editing network. Their intermediate features are fused, and the fused result is blended with the original image to produce the final output.

Model Design

Wrinkle removal is treated as a masked image‑translation task. Three mainstream solutions are considered:

Encoder‑decoder architecture, which combines high‑level semantics from down‑sampling with low‑level details from corresponding up‑sampling layers.

Semantic image editing (e.g., SPADE), which injects global style information from the original image but may lose local details.

StyleGAN‑based methods, which generate high‑quality images but suffer from limited diversity and difficulty preserving identity.

Kuaishou merges the encoder‑decoder structure with semantic editing to retain fine texture while incorporating global style. Four additional techniques are employed: an encoder‑decoder discriminator, RGB‑space skin‑tone correction, GAN‑loss refinement, and frequency‑domain loss.

Discriminator Choice

Four discriminator designs are compared:

Classification GAN – simple binary classifier, loses detail due to multiple down‑samplings.

Patch‑GAN – preserves some detail but is shallow and weak.

No‑down‑sampling discriminator (e.g., ESRGAN) – uses stacked RRDB blocks, retains detail but is computationally heavy.

Encoder‑decoder discriminator – balances high‑level semantics and low‑level detail with moderate cost.

The encoder‑decoder discriminator yields the most realistic skin texture, as shown in Figure 6.

Skin‑Tone Correction

Pixel‑wise L1/L2 loss in RGB space does not correlate well with human perception. To address this, a cosine similarity loss is applied after offsetting low‑intensity RGB values, ensuring that dark regions (where human perception is limited) do not dominate the loss.

GAN‑Loss Adjustment

In inpainting, the “fake” label should reflect that non‑masked regions are real. Therefore, patch‑GAN labels are adjusted per region (Figure 8) and optionally smoothed with Gaussian filtering to better match the effective receptive field.

Frequency Supervision

High‑frequency components often cause texture inconsistency. By supervising the high‑frequency part of the output (Figure 11‑12) while discarding extremely high frequencies, the model improves realism without harming optimization.

Conclusion and Outlook

Kuaishou Y‑Tech’s smart de‑wrinkling system accurately locates facial wrinkles and removes them with impressive realism and performance; the neck‑wrinkle removal feature is already live in the app. Future work will continue to refine algorithmic quality and efficiency, leveraging Kuaishou’s broader computer‑vision expertise to deliver richer user experiences.