Faults play a crucial role in the storage and migration of oil and gas, and accurate identification of fault structures is essential for reliable subsurface interpretation and efficient resource development. In recent years, deep learning–based methods have achieved notable progress in fault recognition. However, owing to the pronounced data distribution discrepancy between synthetic and field seismic data, their generalization ability in field scenarios still requires further improvement. Most existing methods adopt single-stage modeling strategies and tend to rely heavily on idealized fault features learned from synthetic data, resulting in limited spatial continuity and reduced stability under complex field geological conditions. To alleviate these limitations, we propose a two-stage fault recognition method guided by coherence-based prior information to enhance model robustness and applicability in synthetic-to-field scenarios. The proposed method simulates the “rough-to-precise” workflow of manual fault interpretation. In the first stage, coherence attributes serve as prior constraints to guide the generation of an initial representation of the overall fault distribution. In the second stage, the fault probability volume from the first stage acts as a fault-related structural prior for fine-scale modeling within localized fault regions. In addition, auxiliary tasks related to fault recognition are introduced at each stage within a multi-task learning framework, enhancing the model's ability to represent fault spatial continuity and structural consistency. Experimental results show that the performance of different baseline methods varies significantly between synthetic and field seismic data. In comparison, the proposed method exhibits improved stability and generalization across multiple field datasets and delineates fault systems with stronger spatial continuity and more reasonable structures. Overall, the proposed method shows practical value under complex geological conditions and provides a reliable solution for field-oriented fault identification.