The proliferation of space debris, driven by increasingly frequent space activities, poses a severe risk to operational satellites and future missions. Consequently, the surveillance and characterization of space objects have become critical. Current methods for attitude estimation of space objects often rely on target-specific models, such as those with a closed set of facets, and typically employ nonlinear filtering or machine learning techniques. However, these approaches often lack general applicability. This study proposes a more general approach to attitude estimation based on photometric light curves. We define a loss function that quantifies the discrepancy between the measured light curve and a theoretical model parameterized by the attitude parameters. By leveraging automatic differentiation to compute the gradients of this function, we use a gradient descent algorithm to achieve rapid and accurate attitude estimation. The proposed method provides a versatile and efficient framework for attitude estimation, overcoming the target-specificity limitations of existing techniques, with the parameters estimation time reduced to 15 s and the mean squared error lowered to three decimal places. Validation on the third-party database demonstrates that the method exhibits strong generalization capability.