Traditional finite element methods (FEM) can accurately simulate the coupled electromagnetic–thermal fields in wireless power transfer (WPT) systems, but they suffer from high computational costs and slow response times in multi-condition analysis and structural optimization. Therefore, this paper proposes a generative deep learning model based on the conditional diffusion mechanism, termed CoUSD, which takes five-dimensional operating parameters such as coil turns, transmitting current, receiving current, lateral offset and vertical distance as conditional inputs. A unified structural network (UniStructNet) is constructed by integrating residual convolutional blocks and Transformer mechanisms, enabling efficient generation and accurate reconstruction of electromagnetic-thermal fields from Gaussian noise. In simulation-based validation, the model achieves a mean squared error (MSE) of 0.43×10⁻³, a mean absolute percentage error (MAPE) of 2.5%, and a high-frequency MSE (HF-MSE) below 0.017, with a single inference time of only 1.8 seconds—approximately 100 times faster than FEM. In experimental validation at discrete measurement points, the predicted results yield a coefficient of determination (R²) close to 0.9 with respect to the measured data. These results demonstrate that the CoUSD model has significant advantages in high-frequency detail recovery and multiphysics field coupling modeling under complex operating conditions, making it a promising tool for efficient design optimization and rapid performance evaluation of WPT systems.