This paper proposes a dynamic kernel prior (DKP) model for unsupervised blind image super-resolution (BSR). The DKP model enables unsupervised and pre-training-free learning-based algorithms for BSR by adaptively learning dynamic kernel priors through a Markov chain Monte Carlo (MCMC) sampling process on random kernel distributions. The learned kernel prior is then used to optimize a blur kernel estimation network, which employs a network-based Langevin dynamic optimization strategy. These techniques ensure accurate kernel estimation, allowing the DKP model to be easily integrated into existing image restoration models, such as Double-DIP and FKP-DIP, or added to off-the-shelf image restoration models like diffusion models. The DKP model is combined with DIP and diffusion models to validate its performance. Extensive simulations on Gaussian and motion kernel scenarios demonstrate that the DKP model significantly improves kernel estimation with comparable runtime and memory usage, leading to state-of-the-art BSR results. The DKP model is designed to be plug-and-play, requiring no labeled training data or pre-training, and is effective in various kernel scenarios, including Gaussian and motion kernels. The model's flexibility and efficiency make it suitable for real-world applications, offering superior performance in image restoration tasks.This paper proposes a dynamic kernel prior (DKP) model for unsupervised blind image super-resolution (BSR). The DKP model enables unsupervised and pre-training-free learning-based algorithms for BSR by adaptively learning dynamic kernel priors through a Markov chain Monte Carlo (MCMC) sampling process on random kernel distributions. The learned kernel prior is then used to optimize a blur kernel estimation network, which employs a network-based Langevin dynamic optimization strategy. These techniques ensure accurate kernel estimation, allowing the DKP model to be easily integrated into existing image restoration models, such as Double-DIP and FKP-DIP, or added to off-the-shelf image restoration models like diffusion models. The DKP model is combined with DIP and diffusion models to validate its performance. Extensive simulations on Gaussian and motion kernel scenarios demonstrate that the DKP model significantly improves kernel estimation with comparable runtime and memory usage, leading to state-of-the-art BSR results. The DKP model is designed to be plug-and-play, requiring no labeled training data or pre-training, and is effective in various kernel scenarios, including Gaussian and motion kernels. The model's flexibility and efficiency make it suitable for real-world applications, offering superior performance in image restoration tasks.