A diffusion-Mamba synergetic framework for face retouching reversal

Fengchuang Xing

doi:10.1016/j.engappai.2026.115170

A diffusion-Mamba synergetic framework for face retouching reversal

2026-05-22

0

PRE

AI

DOI：10.1016/j.engappai.2026.115170

Original

Original request for help

Save

Abstract

En 中文

Face retouching reversal (FRR) aims to recover authentic facial details from digitally beautified portraits. This task is particularly challenging due to the spatially localized and semantic nature of cosmetic edits. While recent diffusion-based methods demonstrate strong generative capability, most existing approaches lack effective regional structural guidance, often resulting in identity drift or spatial artifacts. To address these limitations, we propose DeltaFRR, which reformulates FRR as a discrepancy-guided reversal process rather than a direct image restoration task. Instead of synthesizing the entire face from scratch, our first stage explicitly models retouching-induced deviations via a pixel-wise Delta-map. This task-aware formulation enables the model to concentrate its generative capacity exclusively on altered regions while preserving unaffected identity anchors. The second stage introduces a Mamba-based enhancement module to restore high-frequency textures while maintaining global semantic consistency across facial components. Extensive experiments on both synthetic and real-world datasets demonstrate that DeltaFRR consistently outperforms state-of-the-art methods in perceptual quality, identity preservation, and cross-dataset generalization.

Keywords:

Face retouching reversal

Diffusion models

Mamba architecture

Delta-map

Identity preservation

A diffusion-Mamba synergetic framework for face retouching reversal

Abstract

Journal

Organization

Cited Papers