David Chung 
118 words
1 minute
Research Note -- Im2Haircut - Single-view Strand-based Hair Reconstruction for Human Avatars
Novelty of this paper
- Uses both real and synthetic data to learn an effective hairstyle prior.
Contributions of this paper:
- Transformer based prior model
- synthetic data for the internal hairstyle geometry
- real data for the outer structure
- Uses transformer based prior to create a Gaussian-splitting-based reconstruction mehtod.
Architecture
- Training the coarse branch considering the synthetic dataset with 3D supervision on the output.
- Training the fine branch with the same losses.
- Joint training with additional real data using self-supervised rendering losses.
Im2Haircut Figure 1
Thoughts
Good parts
- The concept of training synthetic dataset for inner hair structure and real dataset for outer details.
- This paper focus on transformner based prior model while DiffLocks focus on diffusion based model.
Improvements
- The backside of the hair is too smooth.
- The hair color looks the same.
Im2Haircut Figure 2
Research Note -- Im2Haircut - Single-view Strand-based Hair Reconstruction for Human Avatars
http://localhost:4321/posts/research_notes/im2haircut/2025-12-10-im2haircut-notes/ Last updated on 2025-12-10