The diversity of manipulation techniques poses a major challenge for training generalizable image manipulation localization methods. Existing public datasets are often limited in scope and focus on specific manipulation types such as splicing, copy-move, or inpainting, which restricts cross-domain performance. To address this, we introduce GSPA-Net, a novel framework for generating universal training data by simulating realistic manipulation intent, that is, concealing semantic inconsistencies and physical anomalies that localization methods rely on. GSPA-Net employs a dual-disentangled encoder separating manipulation traces into semantic-style and physical-noise representations, which are selectively fused through a domain-aware attention mechanism. This design enables the generated manipulated images to exhibit coherent semantic structures and consistent physical noise distributions, effectively capturing the shared characteristics across diverse manipulation methods. Experiments demonstrate that training localization models on the dataset synthesized by GSPA-Net leads to improved generalization across multiple benchmarks. Our method provides a scalable and practical solution for constructing training data without relying on hand-crafted manipulation pipelines.