Adaptive Blend Pyramid Network for Real-Time Local Retouching of Ultra High-Resolution Images

With the rapid growth of digital media, AI‑driven image editing, especially high‑resolution portrait beautification, has become essential. Traditional filter‑based methods struggle with professional‑grade requirements such as preserving skin texture, handling diverse blemishes, and processing 4K‑8K images within strict time constraints.

Existing deep‑learning approaches—image‑to‑image translation, photo retouching, inpainting, and high‑resolution editing—either lack local focus, require explicit masks, or are computationally heavy, making them unsuitable for fine‑grained, real‑time skin retouching.

To overcome these limitations, the authors propose ABPN, which consists of two main components: a Context‑aware Local Retouching Layer (LRL) that predicts precise skin masks and performs localized edits, and an Adaptive Blend Pyramid Layer (BPL) that progressively upsamples the low‑resolution results while preserving detail through Adaptive Blend Modules (ABM) and their inverse (R‑ABM).

The ABM treats the blend layer as an intermediate representation, enabling adaptive conversion between the original image and the edited result with learnable parameters, while R‑ABM reconstructs the blend layer from low‑resolution outputs. This design reduces unnecessary information, improves efficiency, and maintains high‑fidelity textures.

Extensive experiments on facial skin‑retouching and garment‑wrinkle removal datasets demonstrate that ABPN outperforms state‑of‑the‑art methods in both visual quality and inference speed, achieving real‑time processing of 4K images on a single P100 GPU. Ablation studies confirm the effectiveness of each module, and detailed loss functions are provided to guide training.

Visual comparisons show that ABPN removes blemishes while retaining natural skin texture, and the same framework successfully eliminates garment wrinkles, confirming its versatility for various local editing tasks.

The paper concludes with a comprehensive bibliography of related works in image translation, retouching, and high‑resolution synthesis.