HairPort uses 3D-aware AI to transfer hairstyles between any images
New framework handles large pose gaps with a Bald Converter and 6,000-image dataset
HairPort, a new framework from researchers at Simon Fraser University and collaborators, tackles the long-standing challenge of transferring hairstyles between images where source and target have large viewpoint and scale differences. Unlike prior methods that fail when missing hair must be synthesized, HairPort explicitly separates hair removal from transfer and enforces geometric consistency. The pipeline first uses a Bald Converter — a LoRA-based adaptation of FLUX.1 Kontext — to produce realistic bald versions of input faces. This converter was trained on a new dataset, Baldy, containing 6,000 paired bald and original images spanning diverse identities and conditions.
The transfer then uses a 3D-aware pipeline: it reconstructs the reference hairstyle in 3D and re-renders it from the target viewpoint before compositing onto the bald source image. Finally, a conditional flow-matching generator refines the result. This approach enables accurate, pose-consistent, and identity-preserving hairstyle transfer even under extreme pose and scale discrepancies, significantly outperforming existing methods qualitatively and quantitatively. The work has been accepted to SIGGRAPH 2026 and includes a project page with supplementary materials.
- Uses a Bald Converter with LoRA-based adaptation of FLUX.1 Kontext to generate realistic bald versions of faces.
- Introduces the Baldy dataset with 6,000 paired bald and original images for training the converter.
- 3D-aware pipeline reconstructs reference hairstyles and re-renders from target viewpoint before compositing.
- Conditional flow-matching generator synthesizes final transferred result from bald source and geometry-aligned guidance.
- Accepted to SIGGRAPH 2026 (Conference Papers Track).
Why It Matters
Enables realistic, pose-robust hairstyle transfer for virtual try-ons, AR, and VFX without physical hair changes.