ColorFlow
A three-stage diffusion-based framework, from Tencent ARC, for automatically colorizing black-and-white image sequences, such as cartoon or comic panels, while keeping the same character and object identity across frames.
ColorFlow proposes a Retrieval Augmented Colorization pipeline that colorizes an image using relevant color references, rather than requiring per-character fine-tuning or explicit identity-embedding extraction the way earlier approaches did. A dual-branch design separates color-identity extraction from colorization itself, and both branches use the self-attention mechanism of diffusion models for in-context learning and identity matching.
The authors introduce ColorFlow-Bench, a benchmark for reference-based colorization, and report that ColorFlow outperforms existing models across multiple metrics on it. Inference code and model weights, including a Sketch_Shading model, are released, with training code still to come.
It requires an Anaconda or Miniconda Python environment and is set up by cloning the repository and creating a dedicated conda environment.