ART – Anchoring on Reality: Makeup Transfer

2K Makeup Transfer Visual Showcase

High-Fidelity Makeup Transfer at 2K Resolution

Our framework enables high-fidelity makeup transfer at up to 2K resolution. ART faithfully transfers complex artistic styles and high-frequency cosmetic details while strictly preserving the original source identity and underlying facial geometry.

Drag the center slider to compare. Click side cards or arrows to navigate.

Source

Ref

SourceOurs (2K)

Source

Ref

SourceOurs (2K)

Source

Ref

SourceOurs (2K)

Source

Ref

SourceOurs (2K)

Source

Ref

SourceOurs (2K)

Source

Ref

SourceOurs (2K)

SOTA Comparison

How We Compare Against Prior Work

Select a source–reference pair, then click a method card to compare it against ours in the slider.

Source

Ref

Pair 1

Source

Ref

Pair 2

Source

Ref

Pair 3

Source

Ref

Pair 4

Source

Ref

Pair 5

Source

Ref

Pair 6

SourceOurs

Ours

Nano Banana Pro

GPT Image 1.5

StableMakeup

SHMT

MAD

EleGANt

PSGAN

More Visual Results

Mix & Match Any Source and Reference

Pick any Source and any Reference independently to see different combination effects.

Source

Ref

SourceOurs

No result for this combination

Abstract

What is ART?

"

Existing makeup transfer methods rely on pseudo-target supervision generated by a teacher model, creating a performance ceiling bounded by the teacher's quality. We propose ART (Anchoring on Reality), a two-stage framework that breaks this ceiling by anchoring a refinement cycle directly to real reference images, effectively overriding pseudo-target artifacts.

To support high-fidelity evaluation, we introduce MakeupFaces2K (MF2K), the first 2K-resolution in-the-wild makeup dataset with 8,573 images spanning diverse styles and demographics.

Comprehensive experiments demonstrate state-of-the-art visual fidelity with faithful identity preservation across varying makeup intensities, even under heavy occlusions, extreme expressions, and cross-ethnicity scenarios.

Method Overview

The ART Pipeline

1Reality-Anchored Refinement: By anchoring a refinement cycle to the real reference, our model effectively overrides pseudo-target artifacts while stabilizing training via a controlled-noise bottleneck.

22K-Resolution In-the-Wild Makeup Dataset: We introduce MakeupFaces2K (MF2K), the first 2K-resolution makeup portrait dataset featuring 8,573 images across diverse makeup intensities and underrepresented demographics.

3State-of-the-Art Performance: Our method achieves superior visual fidelity and consistently preserves fine-grained details and source identity, even under occlusions and extreme expressions.

MF2K Dataset

The First 2K Makeup Portrait Dataset

To facilitate research in high-fidelity synthesis, we introduce MakeupFaces2K (MF2K), the first 2K-resolution in-the-wild makeup dataset comprising 8,573 images at 2048×2048 resolution. MF2K covers diverse makeup intensities including bare skin (3,139), light makeup (2,063), heavy makeup (1,798), and artistic styles (1,573), and improves coverage of male portraits and complex high-frequency patterns.

Hover over any sample below to magnify fine-grained cosmetic details.

BibTeX

Cite This Work

@article{wei2026art,
  title={Anchoring on Reality: Breaking the Pseudo-Target Ceiling in Makeup Transfer},
  author={Wei, Bo and Lin, Xianhui and Dong, Yi and Li, Zhongzhong and Li, Zonghui and Wang, Zirui and Yang, Jiachen and Liu, Xing and Gu, Hong and Li, Xiaoming and Zuo, Wangmeng},
  journal={arXiv preprint arXiv:2606.31089},
  year={2026}
}