DiT-based Controllable Hair Style Editing

We build controllable hair-editing systems on top of diffusion-transformer (DiT) backbones. Building on our prior GAN-inversion work for hairstyle manipulation, this line of research moves to modern DiT generative priors and explores precise, sketch-driven, identity-preserving hair editing for realistic portrait imagery.

Recent directions include:

1. Sketch-guided Hair Inpainting: User-provided strokes condition the generation of new hairstyles that respect both the input sketch and the surrounding facial context.
2. Matte-gated ControlNet: A novel control architecture that uses per-pixel hair mattes to gate residual conditioning, enabling high-fidelity editing without leaking content into non-hair regions.
3. Context-robust DiT Generation: Studying how DiT generators interact with surrounding context (face geometry, illumination), and proposing conditioning strategies that yield robust generations across diverse inputs.

Hair editing — bridging earlier GAN-inversion work with our DiT-based pipeline.

programming experience

Python, PyTorch, Diffusers, ControlNet/DiT frameworks