ECCV 2026

MagicMakeup

A region-controllable diffusion transformer for high-fidelity makeup transfer.

Ziyi Wang, Siming Zheng, Yang Yang, Shusong Xu, Hao Zhang, Bo Li, Changqing Zou, Peng-Tao Jiang State Key Lab of CAD&CG, Zhejiang University / vivo BlueImage Lab, vivo Mobile Communication Co., Ltd.
Highlights

Precise makeup transfer without losing identity

MagicMakeup transfers full-face and regional makeup while preserving source identity, facial geometry, and non-edited regions.

Project Overview
MagicMakeup overview
Mix references for eyes and lips, edit a single region, or apply full-face makeup.
01Token-Aligned Region Gating

Pixel masks are aligned with attention tokens, then region-specific logit gating blocks cross-region leakage.

02Cross-Modal Perception Guidance

Text and image features are aligned to clarify which attributes should be transferred and which identity cues should be preserved.

03High-Resolution Benchmark

The benchmark covers synthetic and real 1024 x 1024 makeup-transfer settings with region-level evaluation.

1024High-resolution paired data
3Region modes: eyes, lips, face
2Synthetic and real benchmarks
Results

Full-face and regional transfer results

Full-Face Makeup

Full-face makeup reference
Source face Full-face makeup result
Full-face makeup reference
Source face Full-face makeup result
Full-face makeup reference
Source face Full-face makeup result
Full-face makeup reference
Source face Full-face makeup result
Full-face makeup reference
Source face Full-face makeup result
Full-face makeup reference
Source face Full-face makeup result

Regional Transfer

Reference

Output

Regional transfer output

Reference

Output

Regional transfer output