01
Region-Controllable Framework
A region-controllable DiT framework that achieves state-of-the-art performance across diverse styles and real-world data, with strong robustness and generalization.
ECCV 2026
MagicMakeup: A Region-Controllable Diffusion Transformer for High-Fidelity Makeup-Transfer
MagicMakeup
Identity Preserving✦Region Controllable✦High Resolution✦Makeup Transfer✦
Identity Preserving✦Region Controllable✦High Resolution✦Makeup Transfer✦
01 / Highlights
Region-Controllable
Region-Controllable
High-Fidelity Transfer
MagicMakeup transfers reference makeup while preserving source identity and facial geometry, with precise control over full-face, eyes, and lip regions.
Mix eyes and lip references, edit a single region, or apply full-face makeup within one unified framework.
02
TARG and CMPG Modules
Token-aligned region constraints and transfer-preservation disentanglement enable precise regional transfer with reduced spillover and improved identity consistency.
03
High-Resolution Data and Benchmark
An automated makeup-removal pipeline constructs identity-consistent, region-labeled pairs, while MakeupHQ Bench standardizes evaluation in synthetic and real settings.
02 / Results
Full-Face Makeup Transfer
MagicMakeup preserves identity and scene geometry while transferring fine-grained makeup details across diverse styles, poses, and backgrounds.
Source / ResultSource 01 · Reference 01
↔ Drag to compare
Source
MagicMakeup
More Results
Regional Makeup Transfer
Apply eyes or lip makeup within the specified region while preserving source identity, non-edited areas, and scene geometry.
Eyes
Apply Eyes Makeup
Select a reference and drag the slider
Eyes References
SourceMagicMakeup
Lips
Apply Lip Makeup
Select a reference and drag the slider
Lip References
SourceMagicMakeup
03 / Comparisons
Qualitative Comparisons
Representative GAN- and diffusion-based methods are evaluated under the same settings for direct comparison of identity preservation, makeup fidelity, and spatial stability.
SourceSHMT
SourceMagicMakeup
Baseline Method
Drag either output to compare with the source
04 / Method
Spatial Control and Concept Disentanglement
MagicMakeup jointly processes source, reference, and text conditions within MM-DiT to transfer fine-grained makeup while preserving source identity and facial structure.
TARG and CMPG provide complementary spatial and semantic guidance within MM-DiT.
01
Token-Aligned Region Gating
TARG projects facial ROI masks onto the token grid and applies region-specific logit gating, preventing non-ROI tokens from attending to reference makeup cues.
02
Cross-Modal Perception Guidance
CMPG aligns preservation and transfer concepts with their corresponding image features at each denoising step, reducing identity leakage and attribute drift.
05 / Benchmark
MakeupHQ Benchmark
A unified 1024 × 1024 benchmark spanning MakeupHQ-Synthetic and MakeupHQ-Real supports standardized evaluation of identity preservation, makeup fidelity, unedited-region consistency, and distribution quality.
10242High-resolution evaluation
03Face, eyes, and lip regions
02Synthetic and real test sets
Data Pipeline
Region-specific makeup removal and cascaded quality filtering produce 6,772 identity-consistent, region-labeled pairs across face, eyes, and lip splits.
Our Benchmark
MakeupHQ-Synthetic and MakeupHQ-Real provide image- and identity-disjoint evaluation across diverse styles, poses, expressions, and demographics.
06 / Citation
Cite MagicMakeup
If this project supports your research, please cite the ECCV 2026 paper below.
BibTeX
@misc{wang2026magicmakeupregioncontrollablediffusiontransformer,
title={MagicMakeup: A Region-Controllable Diffusion Transformer for High-Fidelity Makeup-Transfer},
author={Ziyi Wang and Siming Zheng and Yang Yang and Shusong Xu and Hao Zhang and Bo Li and Changqing Zou and Peng-Tao Jiang},
year={2026},
eprint={2607.20924},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2607.20924},
}
}