MakeupMirror: Improving Facial Attribute Preservation in Diffusion Models for Makeup Transfer

A team of researchers has introduced MakeupMirror, a diffusion-based makeup transfer model designed to better preserve facial identity and skin tone, addressing key shortcomings in existing virtual try-on systems [1]. The model, detailed in a paper submitted to arXiv on June 18, 2026, targets limitations in current state-of-the-art solutions such as Stable-Makeup, which struggle with identity and skin color preservation, making them unsuitable for production-level virtual try-on (VTO) in online shopping [1]. MakeupMirror introduces several technical modifications to the underlying diffusion architecture. It integrates facial geometry conditioning using ControlNets to maintain facial fidelity, applies region-specific transfer control for precise application on skin, eyes, and lips, and uses skin tone-based modulation to prevent alteration during cross-subject transfer [1]. A Levenberg-Marquardt Langevin sampler is also incorporated to accelerate inference while preserving generation quality [1]. In experiments across the CPM-Real, Makeup Wild, and a newly collected, more diverse MakeupSelfies dataset, MakeupMirror improved relative facial recognition similarity by +60% and reduced relative skin tone difference by -50% compared to Stable-Makeup [1]. The model achieved an inference latency of 0.7 seconds and an expert acceptance rate of 94% on core facial identity preservation criteria [1]. The paper appears on arXiv, an open-access repository for electronic preprints that, as of late 2024, receives about 24,000 submissions per month and has hosted over two million articles since its founding in 1991 [6]. The repository is not peer-reviewed but serves as a primary distribution channel in fields such as computer science and physics [6]. The MakeupMirror paper is listed under the Computer Vision and Pattern Recognition category and is accompanied by links to experimental community tools available through arXivLabs, a framework for third-party collaborators to build features on the platform [4][5]. These tools include citation explorers and code finders, though they are independent of the paper's peer-review status [5].