With the development of digital art, automated image processing technologies are gaining increasing attention. Recently, a research team from Tsinghua University and Tencent ARC Lab introduced a novel image sequence coloring model named ColorFlow. This model aims to address the challenge of maintaining character and object identity consistency while coloring black-and-white image sequences, catering to the practical needs of industries such as comics and animation.
ColorFlow is a three-stage diffusion-based framework that fully utilizes contextual information to accurately generate colors for black-and-white image sequences through a reference image pool. For instance, the model can effectively color characters' hair and clothing, ensuring color consistency with reference images. Unlike previous techniques that required fine-tuning for each character, ColorFlow simplifies the color generation process through an innovative, retrieval-enhanced coloring pipeline with strong generalization capabilities.
The model's design includes two main branches: one for extracting color identity and the other responsible for the actual coloring process. This dual-branch design leverages the advantages of diffusion models, enabling powerful contextual learning and color identity matching through a self-attention mechanism. To validate the effectiveness of ColorFlow, the research team also launched ColorFlow-Bench, a comprehensive benchmark specifically for reference image-based coloring tasks.
In comparative experiments, ColorFlow outperformed existing advanced models across multiple metrics, demonstrating higher aesthetic quality and generating colors that closely match the original images. The research team showcased ColorFlow's application effects in various artistic scenarios, including black-and-white comics, line art, real-world photographs, and cartoon storyboards, all achieving satisfactory results.
The launch of ColorFlow not only sets a new benchmark for automated coloring technology for image sequences but also provides strong support for further development in the art industry. The research team hopes this technology can achieve wider promotion in practical applications, driving innovation and progress in digital art creation.
Project link: https://zhuang2002.github.io/ColorFlow/
Key Points:
🌟 ColorFlow is an innovative model for coloring black-and-white image sequences that maintains character identity consistency.
🎨 The model features a dual-branch design for color identity extraction and actual coloring, enhancing the effectiveness and efficiency of the coloring process.
🏆 ColorFlow surpasses existing advanced models across multiple metrics, showcasing higher aesthetic quality and practicality.