Image steganography involves the concealment of confidential information within images, rendering it undetectable to unauthorized observers, and subsequently retrieving the hidden information following secure transmission. Numerous investigations into image steganography employ invertible neural networks. Nevertheless, the intricate architecture of these models poses significant challenges to both their training and inference processes. Consequently, this paper proposes a high-capacity image steganography scheme utilizing compressible modules. We propose a multi-scale attention module that compresses the model structure through structural reparameterization post-training, thereby enhancing inference speed. Additionally, we employ a pre-trained image autoencoder to extract deep features from large-scale images, facilitating high-capacity steganography within invertible neural networks by concealing key features. Experimental results indicate that our approach offers superior image quality and model inference speed, while significantly enhancing security relative to existing methodologies.