Cobra - An open-source comic line art coloring framework developed by Tsinghua University, the Chinese University of Hong Kong, and Tencent.

What is Cobra?

Cobra (Efficient Line Art Colorization with BRoAder References) is a comic line art coloring framework developed by Tsinghua University, the Chinese University of Hong Kong, and Tencent ARC Labs. It is designed for high-precision, high-efficiency, and flexible industrial applications. The framework integrates over 200 reference images and uses Causal Sparse Attention and locally reusable positional encoding techniques to effectively manage long-term contextual information, ensuring color consistency and identity preservation. Cobra supports color hints, allowing users to flexibly adjust the coloring results. Cobra's core architecture, Causal Sparse DiT, significantly improves inference speed and interactivity, meeting the demands for contextual consistency and rapid coloring in comic production.

Cobra's main functions

• High-precision line art coloringSupports converting black and white line art into illustrations with rich colors and consistent details.
• Efficient reasoning and interactivityImprove reasoning speed to meet industrial-grade requirements for real-time performance and interactivity.
• Flexible color cues supportUsers can adjust the color of specific areas based on color hints, enhancing the flexibility and personalization of coloring.
• Suitable for a variety of application scenariosSupports extended tasks such as line art with shadows and coloring animated videos.

Cobra's technical principles

• Causal Sparse DiT ArchitectureEliminating pairwise attention computation between reference images significantly reduces computational complexity. Unidirectional causal attention and key-value caching further reduce memory and computational costs. Based on dividing the line art image into multiple local regions and assigning independent positional encoding to each region, any number of reference images can be integrated without altering the pre-trained two-dimensional positional encoding.
• Long context reference managementIt uses a large number of reference images to provide rich color information, and based on the causal sparse attention mechanism, it ensures that the color information of the reference images is efficiently transferred to the target line drawing, avoiding redundant calculations.
• Line Art GuiderThe line art guide integrates the features of the line art image and color cues into the main branch, and achieves precise control over the line art based on the self-attention mechanism, supporting the flexible application of color cues.
• Color-coded integrationThe user specifies the color adjustment result for a particular area. Cobra avoids sampling at edge intersections by limiting the RGB value variation range of the color cue points, reducing ambiguity and improving the accuracy of user control.

Cobra's project address

• Project official website:https://zhuang2002.github.io/Cobra/
• GitHub repository:https://github.com/Zhuang2002/Cobra
• HuggingFace model library:https://huggingface.co/JunhaoZhuang/Cobra
• arXiv technical paper:https://arxiv.org/pdf/2504.12240

Cobra application scenarios

• comic artistQuickly color black and white line art, maintain color consistency between characters and scenes, and improve creative efficiency.
• Animation production teamBatch processing of animation frames for coloring, ensuring color consistency, suitable for animation video production.
• Digital IllustratorGenerates high-quality color illustrations, supports personalized color adjustments, and meets creative needs.
• educatorsAs a teaching tool, it helps students quickly master coloring techniques and enhances learning interactivity.
• Content creatorsQuickly generate images for social media and self-media, enhance content appeal, and meet diverse visual needs.