[16:21 Wed,29.July 2026 � by Thomas Richter] ��� |
With the right prompts, AI-generated videos can nowadays hardly be distinguished from real ones with the naked eye – at the same time, more and more AI clips are flooding the web. Every video is thus potentially suspected of being AI-generated – a major problem for general trust and, of course, especially for the media. As guarantors of "authenticity" amidst the mass of videos posted on social media, the media bear a special responsibility, which also sets them apart from other anonymous sources in a positive way because they are more credible than these.
However, just like any other user on the web, these media also face the problem of distinguishing AI from real when videos do not come through established channels such as news agencies, but from the web, and – especially in the case of current events – can be just as breaking and important as they can be fake. Although portals like  YouTube has introduced automatic AI detection and labeling, this does not provide 100% reliability.  Now, at the ongoing SIGGRAPH 2026 in Los Angeles, NVIDIA – one of the (indirect) main drivers and beneficiaries of the AI boom – has introduced the Synthetic Video Detector, a tool designed to help with precisely this task: It is supposed to detect with up to 92% accuracy whether a clip was created by generators like Veo, Sora, Runway, Kling, or similar ones. To do this, it analyzes individual frames as well as temporal features for typical statistical-characteristic patterns left behind by diffusion models during video generation. As a result of this analysis, it provides a score for the entire video, indicating how high it estimates the probability that the video was synthetically generated. However, this rate – which should be treated with caution – only applies to uncompressed 1080p source material; for heavily compressed videos – i.e., many social media clips – the detection rate drops to 82%. According to NVIDIA, however, the method is supposed to be relatively robust and capable of detecting AI videos even after scaling, cropping, and re-encoding. For now, NVIDIA is offering the Synthetic Video Detector specifically to newsrooms, giving them a signal for the authenticity of a video that complements previous methods. It is available in the form of a NIM (NVIDIA Inference Microservices), meaning an AI model packaged by NVIDIA that can be deployed immediately locally or in one&s own cloud. According to NVIDIA, the processing speed is very high: 22 milliseconds per 1080p video on NVIDIA RTX systems, around 30 ms on L40 GPUs.  While the 82% detection rate is quite good, it naturally also carries the risk of many misclassified videos – nevertheless, the Synthetic Video Detector can help as an automated pre-filter during screening in professional environments with a large input of clips. The NVIDIA Synthetic Video Detector can be tried out online, but is currently not publicly accessible as part of the "AI for Media Private Access Program Service". The first official integration partner is Wowza, a video streaming infrastructure provider active with over 35,000 deployments in more than 170 countries, which intends to embed SVD into its Video Intelligence Framework. AI or not?AI images often have the typical "slop" look, but that doesn&t have to be the case – depending on the source image or prompt, videos (and of course images) can also look completely realistic.  Anyone can test how well AI and non-AI videos can be distinguished by trying out interactive online tests such as AI or Not?, AI Quiz or Real or AI?. How good is your AI detection rate?  �more infos at bei build.nvidia.comdeutsche Version dieser Seite: Nvidias neuer Synthetic Video Detector soll KI-Videos erkennen |
 |
|
|