sanghyeonlee/region-specific-colorization

Enabling Region-Specific Control via Lassos in Point-Based Colorization (AAAI 2025)

This repository provides pretrained checkpoints for the paper Enabling Region-Specific Control via Lassos in Point-Based Colorization.

Paper: https://arxiv.org/abs/2412.13469

Model Details

• —Architecture: Vision Transformer with localized attention masks guided by user-provided lasso regions.
• —Interaction: Point-based color hints + lasso regions to constrain propagation.
• —Framework: PyTorch

Checkpoints

• —icoloritv2lab_base_patch16_224_henc6_patchloss.pth (recommended)
• —icoloritv2lab_base_patch16_224_henc6_fixmask.pth

Intended Use

Interactive image colorization from grayscale inputs with user-provided color hints and optional lasso constraints.

Limitations

• —Performance depends on the quality and placement of user hints/lassos.
• —May produce artifacts on extremely low-resolution inputs or unusual textures.

Citation

@inproceedings{lee2025enabling,
  title={Enabling Region-Specific Control via Lassos in Point-Based Colorization},
  author={Lee, Sanghyeon and Yun, Jooyeol and Choo, Jaegul},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={5},
  pages={4544--4552},
  year={2025}
}