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

MakeupMirror is a diffusion-based makeup transfer method aimed at fixing the two problems that make these systems hard to ship: identity drift and unintended skin-tone changes. The authors position it as a step beyond Stable-Makeup, especially for virtual try-on and AR-style cosmetic previews where preserving the person underneath matters as much as the makeup itself.

The paper adds four pieces that matter in practice: facial geometry conditioning via ControlNets, region-specific control for skin/eyes/lips, skin-tone-aware modulation for cross-subject transfer, and a Levenberg-Marquardt Langevin sampler to cut latency. That last part is notable because the reported inference time is 0.7s, which is much closer to something you could actually prototype into a consumer-facing flow than many diffusion demos.

For game developers, the relevance is mostly in character customization, avatar creation, and any pipeline that uses face beautification or cosmetic overlays. If you’ve ever fought a model that “improves” a face by subtly changing who the character is, the facial-fidelity angle here is the real story. The region-based control could also be useful for tools teams building artist-facing makeup or facial detail editors.

Results are reported on CPM-Real, Makeup Wild, and a newly collected MakeupSelfies dataset. The paper claims +60% relative facial recognition similarity, -50% relative skin-tone difference, and a 94% expert acceptance rate on core identity-preservation criteria. It’s still arXiv research, but the combination of better fidelity, better control, and lower...