Face image quality is an important factor to enable high performance facerecognition systems. Face quality assessment aims at estimating the suitabilityof a face image for recognition. Previous work proposed supervised solutionsthat require artificially or human labelled quality values. However, bothlabelling mechanisms are error-prone as they do not rely on a clear definitionof quality and may not know the best characteristics for the utilized facerecognition system. Avoiding the use of inaccurate quality labels, we proposeda novel concept to measure face quality based on an arbitrary face recognitionmodel. By determining the embedding variations generated from randomsubnetworks of a face model, the robustness of a sample representation andthus, its quality is estimated. The experiments are conducted in across-database evaluation setting on three publicly available databases. Wecompare our proposed solution on two face embeddings against sixstate-of-the-art approaches from academia and industry. The results show thatour unsupervised solution outperforms all other approaches in the majority ofthe investigated scenarios. In contrast to previous works, the proposedsolution shows a stable performance over all scenarios. Utilizing the deployedface recognition model for our face quality assessment methodology avoids thetraining phase completely and further outperforms all baseline approaches by alarge margin. Our solution can be easily integrated into current facerecognition systems and can be modified to other tasks beyond face recognition.