Research map / Image generation / Editing and style transfer
What to Preserve and What to Transfer: Faithful, Identity-Preserving Diffusion-based Hairstyle Transfer
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 HairFusion, a one-stage diffusion-based hairstyle transfer model that uses a hair-agnostic input, pose-aware hair align cross-attention, and adaptive hair blending to transfer hairstyles on in-the-wild face images.
Problem
Hairstyle transfer that modifies a face image's hair to a reference hairstyle while preserving identity, clothing, and background; existing StyleGAN-based methods fail to generalize to extreme head poses and focal lengths and lose fine hair detail.
Approach
Reframes hairstyle transfer as exemplar-based inpainting with a diffusion model: a hair-agnostic input removes original hair, a pose-conditioned Align-CA aligns the reference hair to the face, and adaptive hair blending uses cross-attention-derived hair masks to keep non-hair regions of the source during denoising.
Key results
- On K-Hairstyle, HairFusion reaches FID 10.82 versus 13.26 for Paint-by-Example and 22.84 for Style Your Hair in the transfer setting, with PSNR 21.15 and LPIPS 0.18.
- On cropped and aligned CelebV-Text, FID is 22.00 versus 35.69 for Style Your Hair and 35.40 for HairCLIPv2.
- Ablation on K-Hairstyle: adding adaptive hair blending lowers FID from 15.59 (with Align-CA) to 10.82 and raises PSNR from 19.33 to 21.15.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Align-CA to the baseline (ablation) | HairFusion · K-Hairstyle | SSIM: 0.58 -> 0.62 | Helped |
| Add Align-CA to the baseline (ablation) | HairFusion · K-Hairstyle | PSNR: 18.09 -> 19.33 | Helped |
| Add adaptive hair blending to the model with Align-CA (ablation) | HairFusion · K-Hairstyle | FID: 15.59 -> 10.82 | Helped |
| Add adaptive hair blending to the model with Align-CA (ablation) | HairFusion · K-Hairstyle | SSIM: 0.62 -> 0.70 | Helped |
| Add adaptive hair blending to the model with Align-CA (ablation) | HairFusion · K-Hairstyle | PSNR: 19.33 -> 21.15 | Helped |
| Add adaptive hair blending to the model with Align-CA (ablation) | HairFusion · K-Hairstyle | LPIPS: 0.27 -> 0.18 | Helped |
Surprising findings
- Align-CA alone does not improve FID (15.43 baseline vs 15.59), and the gains in FID come mainly from adaptive hair blending.
- On CelebV-Text reconstruction, Style Your Hair attains slightly higher PSNR (22.09) than HairFusion (21.77) on cropped and aligned images.
Limitations
- Output quality depends on external face parsing and facial landmark detectors, which can miss fine hair strands.
- Struggles to generate or reconstruct detailed features such as text near the hair region.
- Does not support user control over the length or shape of the reference hairstyle.
Experimental setup
Datasets: K-Hairstyle CelebV-Text
Benchmarks: K-Hairstyle CelebV-Text
Metrics: FID SSIM PSNR LPIPS
Baselines: Style Your Hair HairCLIPv2 Paint-by-Example Anydoor HairFastGAN Stable-Hair
Builds on: Stable Diffusion Paint-by-Example Blended Diffusion ControlNet
Closest papers by meaning
- Stable-Hair: Real-World Hair Transfer via Diffusion Model (AAAI 2025 · similarity 0.93)
- HairDiffusion: Vivid Multi-Colored Hair Editing via Latent Diffusion (NeurIPS 2024 · similarity 0.89)
- HairFastGAN: Realistic and Robust Hair Transfer with a Fast Encoder-Based Approach (NeurIPS 2024 · similarity 0.89)
- DiffLocks: Generating 3D Hair from a Single Image using Diffusion Models (CVPR 2025 · similarity 0.88)
- Zero-Shot Head Swapping in Real-World Scenarios (CVPR 2025 · similarity 0.87)
- MonoHair: High-Fidelity Hair Modeling from a Monocular Video (CVPR 2024 · similarity 0.87)
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.