The paper presents a learning-based approach to low-level vision problems, focusing on estimating the underlying scene from image data. The authors use a Markov network to model the statistical dependencies between image patches and scene patches. A factorization approximation simplifies the learning and inference processes, allowing for efficient propagation of image information. Monte Carlo simulations validate the approximation's accuracy. The method is applied to three problems: super-resolution (estimating high-frequency details from low-resolution images), motion estimation (resolving the aperture problem and filling-in), and joint Gaussian processes. The results show good performance, with the method outperforming traditional methods in super-resolution and resolving issues in motion estimation. The technique is flexible and can be applied to other vision problems, such as line drawing interpretation and distinguishing shading from reflectance.The paper presents a learning-based approach to low-level vision problems, focusing on estimating the underlying scene from image data. The authors use a Markov network to model the statistical dependencies between image patches and scene patches. A factorization approximation simplifies the learning and inference processes, allowing for efficient propagation of image information. Monte Carlo simulations validate the approximation's accuracy. The method is applied to three problems: super-resolution (estimating high-frequency details from low-resolution images), motion estimation (resolving the aperture problem and filling-in), and joint Gaussian processes. The results show good performance, with the method outperforming traditional methods in super-resolution and resolving issues in motion estimation. The technique is flexible and can be applied to other vision problems, such as line drawing interpretation and distinguishing shading from reflectance.