FLUX-Makeup: High-Consistency Makeup Transfer Without Extra Face-Control Modules

360 AI Research 2026-01-27 801 views

A simpler makeup-transfer pipeline

Figure 1 from the original article

FLUX-Makeup accepts a source portrait and a makeup reference, then transfers the desired appearance without requiring additional face-control modules. The aim is to preserve the source identity and facial structure while reproducing makeup details faithfully.

Figure 2 from the original article

Figure 3 from the original article

Building a high-quality training set

The project introduces HQMT, a data pipeline that combines language-model-generated makeup descriptions with FLUX-Kontext generation. Starting from roughly 350,000 candidate pairs built on FFHQ images, strict filtering removes identity shifts, excessive stylization, failed transfers, and inconsistent backgrounds. Fewer than 20 percent of candidates pass the full quality process, producing more than 50,000 curated pairs.

Figure 4 from the original article

Figure 5 from the original article

Decoupling identity and makeup

Figure 6 from the original article

The model uses a decoupled injection design so that source identity and reference makeup can be represented separately. This improves makeup fidelity, identity consistency, and robustness across different references and source faces.

Figure 7 from the original article

Code, model weights, and the workflow have been released for reproducible evaluation and further research.

Figure 8 from the original article

Figure 9 from the original article

Back to Tech Blog