Research map / Image generation / Editing and style transfer
Content-Style Decoupling for Unsupervised Makeup Transfer without Generating Pseudo Ground Truth
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CSD-MT, a purely unsupervised makeup transfer method that decouples content (high-frequency) and makeup style (low-frequency) via frequency decomposition, avoiding pseudo ground truth synthesis.
Problem
Makeup transfer lacks real transferred ground truth; existing methods rely on synthesized pseudo ground truths (PGTs) that are sub-optimal and degrade performance.
Approach
Assume low-frequency components of a face carry makeup style and high-frequency components carry content; train a generator by maximizing content consistency (HF with source) and makeup consistency (LF with reference) without any PGT, plus a self-augmented reconstruction loss and a VGG-Gram color contrastive loss.
Key results
- CSD-MT achieves FID 37.56 on the MT dataset, the lowest among compared methods (best baseline SCGAN 39.20 and SSAT 38.01).
- Self-augmented PSNR/SSIM on MT (crop) reaches 27.28/0.920, above EleGANt's 25.15/0.885.
- In a 23-participant user study, CSD-MT was selected as best in 55.9% (simple), 70.4% (complex), and 76.7% (extreme) makeup cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace PGT-guided training with the unsupervised CSD-MT strategy (ablation) | CSD-MT · MT | Self-Aug PSNR (Crop): 23.23 -> 27.28 | Helped |
| replace PGT-guided training with the unsupervised CSD-MT strategy (ablation) | CSD-MT · MT | Self-Aug PSNR (Crop): 22.22 -> 27.28 | Helped |
| add Lmakeup to Ladv + Lcont (ablation) | CSD-MT · MT | Self-Aug PSNR (Crop): 19.97 -> 23.01 | Helped |
| add Lcycle (ablation) | CSD-MT · MT | Self-Aug PSNR (Crop): 23.01 -> 24.92 | Helped |
| add Laug (ablation) | CSD-MT · MT | Self-Aug PSNR (Crop): 24.92 -> 26.45 | Helped |
| add Lcts (ablation) | CSD-MT · MT | Self-Aug PSNR (Crop): 26.45 -> 27.28 | Helped |
| add Laug and Lcts (ablation) | PSGAN · MT | Self-Aug PSNR (Crop): +0.29 | No clear effect |
| add Laug and Lcts (ablation) | SSAT · MT | Self-Aug PSNR (Crop): +0.97 | No clear effect |
| use unsupervised CSD-MT instead of SSAT | makeup transfer methods · MT | FID: 38.01 -> 37.56 | Helped |
| use unsupervised CSD-MT instead of SSAT | makeup transfer methods · Wild-MT | FID: 70.53 -> 60.82 | Helped |
| use unsupervised CSD-MT instead of SpMT | makeup transfer methods · LADN | FID: 48.18 -> 40.87 | Helped |
| use CSD-MT instead of EleGANt | makeup transfer methods · MT | Self-Aug PSNR (Crop): 25.15 -> 27.28 | Helped |
Surprising findings
- Training the same architecture with PGTs from PSGAN or SSAT yields outputs resembling those methods, suggesting PGT quality dominates transfer performance over network structure.
Limitations
- The high-frequency-equals-content assumption may fail for images with patterns.
- Evaluation of transfer quality relies on proxy self-augmented reconstruction metrics since no real ground truth exists.
- User study relies on subjective category labels for simple/complex/extreme makeup.
Experimental setup
Datasets: MT dataset Wild-MT dataset LADN dataset
Benchmarks: MT dataset Wild-MT LADN dataset
Metrics: FID self-augmented PSNR self-augmented SSIM user preference rate
Baselines: BeautyGAN PSGAN SCGAN SpMT LADN SSAT EleGANt
Builds on: CycleGAN LSGAN SPADE PSGAN SSAT
Closest papers by meaning
- SHMT: Self-supervised Hierarchical Makeup Transfer via Latent Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Makeup Prior Models for 3D Facial Makeup Estimation and Applications (CVPR 2024 · similarity 0.85)
- Less is More: Masking Elements in Image Condition Features Avoids Content Leakages in Style Transfer Diffusion Models (ICLR 2025 · similarity 0.84)
- WaveFace: Authentic Face Restoration with Efficient Frequency Recovery (CVPR 2024 · similarity 0.84)
- CLeaR: A Unified Framework for Resolving the Leakage–Degradation Dilemma in Style Transfer (NeurIPS 2026 · similarity 0.84)
- What to Preserve and What to Transfer: Faithful, Identity-Preserving Diffusion-based Hairstyle Transfer (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.