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

The Challenge of Identity Preservation in Makeup Transfer

A new paper, "MakeupMirror," tackles a persistent problem in generative AI: how to transfer makeup between faces without altering the underlying facial identity. While diffusion models have made impressive strides in image-to-image translation, applying a makeup look from a reference image to a target face often introduces unwanted changes to facial features, expressions, or skin texture. The MakeupMirror approach specifically addresses this "attribute preservation" gap, aiming to keep the person recognizable while faithfully applying the cosmetic style.

Why This Matters Beyond Virtual Try-On

The immediate application is obvious: better virtual try-on for cosmetics and more realistic AR filters. But the underlying challenge—preserving identity while modifying appearance—has broader implications. Many generative tasks require this balance: age progression, hairstyle changes, or even medical imaging where pathological features must be highlighted without distorting anatomy. MakeupMirror’s technical approach could inform solutions for any domain where a model must apply a "style" without erasing the subject's core characteristics.

For the beauty industry specifically, this matters because current solutions often produce results that look "off" in subtle ways—eyes slightly reshaped, jawlines softened, or skin texture homogenized. These artifacts break the illusion of a realistic preview, reducing consumer trust in virtual try-on tools. A more robust solution could accelerate adoption of AR makeup shopping, which has remained a novelty rather than a primary purchase driver.

Implications for AI Practitioners

The paper highlights a fundamental tension in diffusion models: the same denoising process that enables high-quality generation also tends to wash out fine-grained identity cues. Practitioners working with diffusion models for any face-related task should note three key lessons:

First, conditioning alone is insufficient. Simply feeding an identity image as a condition doesn't guarantee preservation—the model needs explicit architectural mechanisms to separate identity from style. MakeupMirror likely introduces such separation, which is a pattern worth studying for other tasks.

Second, evaluation metrics for identity preservation remain immature. Most papers report FID or LPIPS scores, but these don't capture whether a face remains recognizable. Practitioners should demand identity-specific metrics (e.g., face verification similarity scores) when evaluating models for face-related tasks.

Third, the trade-off between fidelity and editability is real. Stronger identity preservation may limit how dramatically makeup can be changed. Finding the right balance requires careful tuning, and no single solution works for all use cases—AR filters may tolerate more identity drift than e-commerce try-on.

Key Takeaways

• MakeupMirror addresses a critical failure mode in diffusion-based makeup transfer: unwanted alteration of facial identity during style application
• The problem extends beyond cosmetics to any generative task requiring style transfer without subject distortion
• Practitioners should adopt identity-specific evaluation metrics and architectural separation of identity/style features
• The fidelity-editability trade-off means no universal solution exists—application-specific tuning remains necessary