HairPort: In-context 3D-aware Hair Import and Transfer for Images

HairPort is a new hairstyle transfer pipeline aimed at the cases that usually break image-to-image hair editing: large pose gaps, scale changes, and viewpoint mismatch. Instead of trying to directly paste hair across images, it splits the problem into hair removal, geometry alignment, and final synthesis, which is a much more practical fit for production-style workflows.

The first stage is a Bald Converter built on FLUX.1 Kontext with LoRA-based in-context adaptation. To train it, the team created Baldy, a dataset of 6,000 paired bald and original images spanning different identities and conditions. That gives the system a cleaner starting point than traditional transfer methods, which often struggle when the target face still contains the original hairstyle.

The second stage reconstructs the reference hairstyle in 3D, re-renders it from the target viewpoint, and composites it onto the source image before a conditional flow-matching generator produces the final result. That 3D-aware step is the key technical shift here: it lets the system handle bigger camera and scale differences while keeping the transferred hair aligned with the face.

For game teams, the obvious relevance is in character customization, avatar tools, virtual try-on, and previsualization workflows where hair is one of the hardest assets to edit convincingly from a single image. The practical takeaway is that hair transfer is moving away from brittle 2D blending toward geometry-aware synthesis, which is the direction most real-time and offline character pipelines eventually need anyway.