Generative steganography embeds secret messages during the image synthesis process, producing artificial images that lack the distortion signatures exploited by classical steganalysis. This shift introduces substantial challenges for reliable detection, particularly when the detector encounters stego images generated by previously unseen models. To address these limitations, we propose WaReCo, a wavelet residue contrastive learning framework designed to improve both feature sensitivity and cross-model generalization in deep steganalysis. Our method employs multi-level discrete wavelet transform to decompose images into frequency subbands, capturing high-frequency residual anomalies introduced by embedding operations. Supervised contrastive learning enhances feature discriminability by maximizing intra-class similarity and inter-class separation without requiring traditional cover-stego pairs. Extensive experiments on typical generative steganography methods across multiple datasets demonstrate that WaReCo consistently outperforms existing steganalysis networks. In particular, it achieves substantial gains in cross-model detection accuracy, often improving performance by more than ten percentage points over prior methods when training and testing generative models are different, validating WaReCo’s practical applicability for real-world generative steganalysis.