Cobra – A comic line art coloring framework open-sourced by Tsinghua, The Chinese University of Hong Kong and Tencent

What is Cobra?

Cobra (Efficient Line Art COlorization with BRoAder References) is a manga line art colorization framework developed by Tsinghua University, The Chinese University of Hong Kong, and Tencent ARC Lab. It is designed for high precision, efficiency, and flexibility in industrial applications. The framework integrates over 200 reference images and utilizes Causal Sparse Attention and Locally Reusable Positional Encoding to effectively manage long-range contextual information, ensuring color consistency and identity preservation. Cobra also supports color hints, allowing users to flexibly adjust the coloring results. Its core architecture, Causal Sparse DiT, significantly boosts inference speed and interactivity, meeting the demands for context consistency and fast coloring in manga production.

Key Features of Cobra

High-Precision Line Art Colorization: Converts black-and-white sketches into richly colored illustrations with consistent details.
Efficient Inference and Interactivity: Enhances inference speed to meet the real-time and interactive needs of industrial-grade applications.
Flexible Support for Color Hints: Allows users to fine-tune specific regions using color hints, enhancing the flexibility and personalization of coloring.
Versatile Application Scenarios: Can be extended to tasks such as shadowed line art and animation frame colorization.

Technical Principles of Cobra

Causal Sparse DiT Architecture: Eliminates pairwise attention calculations between reference images, significantly reducing computational complexity. Uses unidirectional causal attention and key-value caching to further reduce memory and computation costs. It divides the line art into multiple local regions and assigns independent positional encodings to each region, integrating any number of reference images without altering pretrained 2D positional encodings.
Long-Context Reference Management: Incorporates a large number of reference images to provide rich color information. Through causal sparse attention, it efficiently transfers color information from references to the target line art, avoiding redundant computation.
Line Art Guider: Integrates features from line art images and color hints into the main branch. Using self-attention mechanisms, it enables precise control of the line art and flexible application of color hints.
Color Hint Integration: Allows users to specify colors for certain regions. Cobra restricts the RGB variation range of hint points and avoids sampling at edge intersections, reducing ambiguity and improving the accuracy of user control.

Cobra Project Links

Project Website: https://zhuang2002.github.io/Cobra/
GitHub Repository: https://github.com/Zhuang2002/Cobra
HuggingFace Model Hub: https://huggingface.co/JunhaoZhuang/Cobra
arXiv Technical Paper: https://arxiv.org/pdf/2504.12240

Application Scenarios of Cobra

Manga Artists: Rapidly colorize black-and-white sketches while maintaining color consistency across characters and scenes, boosting creative efficiency.
Animation Production Teams: Batch-process animation frame colorization with consistent coloring, ideal for animated video production.
Digital Illustrators: Generate high-quality colored illustrations with personalized color adjustments to meet artistic needs.
Educators: Serve as a teaching tool to help students quickly learn coloring techniques and enhance interactive learning.
Content Creators: Quickly generate visuals for social media and multimedia platforms, increasing content appeal and meeting diverse visual requirements.