Institut für Nachrichtentechnik
Univ.-Prof. Dr.-Ing. Alexander Raake
Julius Prenzel, Rakesh Rao Ramachandra Rao, Tom Linke, Christian Rohlfing, Alexander Raake
The rapid progress of text-to-video generation models is making it increasingly hard to detect the synthetic origin of AI-generated videos (AIGVs). At the same time, perceptual evaluation of AIGVs is complex, as no pristine reference exists, and perceived quality depends on multiple interacting factors. One interaction that has already been shown for single images is between generative AI artifacts and distortions caused by compression. We investigate how compression influences both perceived visual quality and authenticity of AIGVs compared to real-world recordings. A dataset of 200 videos (150 AIGVs generated by five state-of-the-art text-to-video models and 50 semantically matched real videos) was encoded using H.264/AVC at five QP levels. The AIGVs were classified into three generative quality levels based on a pretest. In the subjective study, participants rated visual quality and authenticity. The results demonstrate that compression significantly affects authenticity judgments in a content-dependent manner. While lower bit rates decrease the probability of real videos being judged as authentic, medium-quality AIGVs become more likely to be perceived as authentic at lower bit rates, indicating the possibility that stronger compression can partially mask synthetic artifacts. In terms of visual quality, real videos exhibit a monotonic degradation with increasing compression. For high Quantization Parameter (QP), AIGVs show a similar behavior, however for low QP there is no visible impact in quality, suggesting an interaction between generative and compression artifacts. Instrumental quality metrics correlate only moderately with subjective ratings. Overall, the findings reveal that compression potentially reduces visual fidelity and also impacts perceived authenticity of AI-generated videos.
Additional data and plots can be found in this git repository.