Joint intensity–gradient guided generative modeling for colorization

This paper proposes an iterative score-based generative model for solving the automatic colorization problem. Although unsupervised learning methods have shown capability to generate plausible color, inadequate exploration of detailed information and data dimensions still limit performance model. Considering that number samples in has influence on estimating target gradients gradient map possesses important latent image, inference process modeling is conducted joint intensity–gradient domain colorization. Specifically, a set formed high-dimensional tensors are trained, via score matching, attain distribution domain. As function determined, generated by means annealed Langevin dynamics, forming procedure. Furthermore, constraint data-fidelity term proposed degree freedom within at stage, thus being conducive edge-preserving effect. Experimental results conveyed remarkable diversity our method.