This paper proposed kinematic-prior-guided generation of accurate and smooth turning trajectory for autonomous vehicles (AVs). While state-of-the-art methods generate accurate trajectory for car-following and lane-changing, coupled dynamics of car-following and lane-changing in curved turns remain challenging for AVs due to kinematic complexity. To generate accurate and smooth turning trajectory for AVs, this paper proposed Diffusion-Phy. The Diffusion-Phy integrates two novel innovations: (1) Modeling turning kinematics with enhanced physical consistency, and (2) Generating physical anchor points under strong kinematic conditions to guide trajectory generation. The first cascade incorporates three mechanisms: modeling vehicle turning dynamics via differential equations, dynamically correcting states using a dual flow, and projecting this state representation onto multimodal latent features via an implicit state module (ISM) to preserve the motion semantics. The second generates physical anchor points by fusing multimodal inputs with state variables carrying physical semantic features, and constructs an anchored Gaussian distribution with its distribution mean. Conditional sampling is performed in the distribution space by kinematic prior guidance, thereby significantly improving the physical consistency, accuracy, and smoothness of the trajectory in turning scenarios. The Diffusion-Phy outperforms baselines with superior accuracy and smoothness across turn left and turn right scenarios, consistently excelling in both reactive and non-reactive modes.