Research map / Image generation / Editing and style transfer
HairFastGAN: Realistic and Robust Hair Transfer with a Fast Encoder-Based Approach
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 HairFast, an encoder-only hairstyle and color transfer pipeline operating in StyleGAN's FS space with pose alignment, mixing, and refinement modules, achieving near real-time inference.
Problem
Hairstyle transfer from a reference photo to an input face is slow when optimization-based (Barbershop, StyleYourHair, HairCLIPv2) and low quality when encoder-based (HairCLIP in W+), and both struggle with large pose differences between source and reference.
Approach
Decomposes hair transfer into pose alignment (Rotate Encoder in W+ plus shape encoder/adaptor for masks), shape alignment (F-space assembly with SEAN inpainting and E4E regularization), color alignment (S-space modulation encoder with CLIP embeddings and a CLIP cosine loss), and a refinement FS encoder at 64x64 for detail recovery, all using feed-forward encoders with no test-time optimization.
Key results
- HairFast achieves full-transfer FID 13.12 and FID_CLIP 5.12, better than Barbershop (15.94 / 7.07) and CtrlHair (15.10 / 9.52).
- Reconstruction LPIPS 0.08 and PSNR 23.45, outperforming Barbershop (0.11 / 21.18).
- Runs in about 0.78 s on V100 for the full transfer mode, versus 645 s for Barbershop and 7.87 s for CtrlHair.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HairFast instead of HairCLIPv2 | Hair transfer · CelebA-HQ hair transfer (full) | FID: 14.28 -> 13.12 | Helped |
| Use HairFast instead of Barbershop | Hair transfer · CelebA-HQ hair transfer (full) | FID_CLIP: 7.07 -> 5.12 | Helped |
| Use HairFast instead of HairCLIP | Hair transfer · Hair transfer runtime (V100) | time (s): 0.36 -> 0.78 | Hurt |
| Add Color alignment (ablation) | HairFast · CelebA-HQ hair transfer | FID: 26.88 -> 16.23 | Helped |
| Add FS and W+ F-space mixing (ablation) | HairFast · CelebA-HQ hair transfer | FID: 16.57 -> 16.23 | No clear effect |
| Add Rotate Encoder (ablation) | HairFast · CelebA-HQ hair transfer | FID: 16.87 -> 16.23 | Helped |
| Add Shape Adaptor (ablation) | HairFast · CelebA-HQ hair transfer | FID: 18.72 -> 16.23 | Helped |
| Use SEEN inpaint instead of the Barbershop optimization replacement (ablation) | HairFast · CelebA-HQ hair transfer | FID: 12.79 -> 16.23 | Hurt |
| Add Refinement alignment (ablation) | HairFast · CelebA-HQ hair transfer | FID: 16.23 -> 13.12 | Helped |
Surprising findings
- CtrlHair achieves better FID than optimization-based methods like Barbershop and StyleYourHair, yet visually shows artifacts from Poisson blending; the authors argue FID_CLIP is more reliable.
Limitations
- Struggles when long hair is replaced by short hair, with inpainting artifacts.
- Does not handle complex hair textures such as ponytails, ribbons, and braids.
- Color transfer fails under large illumination differences.
Experimental setup
Datasets: FFHQ CelebA-HQ
Benchmarks: CelebA-HQ
Metrics: FID FID_CLIP LPIPS PSNR runtime
Baselines: HairCLIP HairCLIPv2 CtrlHair StyleYourHair Barbershop
Builds on: Barbershop CtrlHair HairCLIP StyleGAN E4E SEAN
Closest papers by meaning
- Stable-Hair: Real-World Hair Transfer via Diffusion Model (AAAI 2025 · similarity 0.89)
- What to Preserve and What to Transfer: Faithful, Identity-Preserving Diffusion-based Hairstyle Transfer (AAAI 2025 · similarity 0.89)
- DGH: Dynamic Gaussian Hair (NeurIPS 2025 · similarity 0.87)
- Perm: A Parametric Representation for Multi-Style 3D Hair Modeling (ICLR 2025 · similarity 0.87)
- MonoHair: High-Fidelity Hair Modeling from a Monocular Video (CVPR 2024 · similarity 0.87)
- DiffLocks: Generating 3D Hair from a Single Image using Diffusion Models (CVPR 2025 · similarity 0.86)
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.