Generative Adversarial Networks (GANs) have revolutionized the generation of synthetic malware images, providing significant applications in cybersecurity. Traditional image-based metrics such as Inception Score (IS) and Fréchet Inception Distance (FID) evaluate the visual quality of these images but fail to capture their malicious nature fully. We propose a novel evaluation framework, cTIMS (Correlated Textual-Based and Image-Based Metric Suites), which integrates both image and text-based metrics to provide a comprehensive assessment. Utilizing the multimodal image captioning model BLIP, we extract textual descriptions of GAN-synthesized malware images and analyze the correlation between image metrics (IS, FID) and text metrics (BLEU, METEOR, ROUGE). Additionally, we introduce Kolmogrov-Arnold Network (KAN)-based CNN architectures (VGG16-KAN, VGG19-KAN), demonstrating significant performance improvements in malware classification. The effectiveness of the Train Real Test Synthetic (TRTS) approach in validating these correlations is also evaluated, confirming the method’s reliability in selecting high-quality GAN-generated images.