Scientific images are essential in many experimental sciences, but the large volumes of data present significant challenges. Effective image compression must be fast, achieve high compression ratios, and preserve important domain-specific features. Existing compressors, such as JPEG or SZ, can distort critical textures at high compression ratios. AI-based compressors, on the other hand, offer excellent image quality and high compression ratios but are significantly slower than traditional methods. To address this discrepancy, we designed \thiswork{}, a high-performance AI-based compressor that preserves \emph{vi}sual \emph{sem}antics. Our method enhances AI compression by incorporating sparse encoding with varied-length integer truncation, optimized lossless encoding using bitshuffle and decoupled lookback prefix-sum, and pipelining for efficient data streaming and asynchronous processing. Evaluations on general and scientific datasets demonstrate that, under similar compression ratios, \thiswork{} preserves the best image quality while achieving a $1.9\times$ speedup (vs. nvJPEG) compared to non-AI-based compressors, and performs almost on par with AI-based compressors while delivering a 9.6$\times$ overall compression speedup. This effectively bridges the performance gap between traditional and AI-based compression methods.