In industrial food processing, significant financial losses occur due to the approval of substandard or defective potatoes that do not meet quality standards. The challenge is that potatoes vary widely in color, shape, and size, and the image data involved is often complex and nonlinear, making anomaly detection difficult. To address these issues, we explore the use of Glow-based Invertible neural networks as Anomaliedetection (INNs) for anomaly detection. While U-NET Autoencoders (UAEs) are effective at capturing complex structures, Glow-based INNs offer advantages in precisely modeling data distributions. Our approach optimizes Glow-based INNs specifically for potato image analysis, leading to improved performance over traditional methods. The enhanced interpretability of Glow-based INNs also supports their integration into industrial settings.