<p>Deep generative models are widely used to generate high-dimensional data such as images. Because these models often do not provide explicit likelihoods, their performance is typically evaluated by comparing samples from real and generated data. The most common metric is the Fréchet Inception distance (FID), which compares the first two moments of embedded features extracted by a pre-trained Inception network. Although FID is popular for its simplicity, many studies have pointed out its limitations when used as the sole evaluation metric. In this work, we propose using the sliced Wasserstein distance (SWD) with CLIP embeddings as a more reliable alternative. We design a controlled synthetic experimental framework to compare image-generation models and show that SWD outperforms FID in detecting small differences between image distributions. We also evaluate the maximum mean discrepancy (MMD) and highlight its practical limitations. Our findings provide practical guidance on selecting appropriate evaluation metrics for image-generation tasks.</p>