[12:04 Fri,14.August 2026 � by Thomas Richter] ��� |
The question of whether a video online is real or AI-generated is becoming increasingly difficult to answer given the rise of evermore realistic AI videos. Researchers from the University of California Riverside, together with YouTube and Google DeepMind, have now introduced a method called SAGA (Source Attribution of Generative AI Videos). It is designed not only to detect AI videos, but also to identify which AI model was used to create them.  SAGA
To do this, SAGA examines videos for characteristic traces that various AI video generators leave behind during image generation. The researchers compare these to a generator's fingerprints. SAGA operates in multiple levels of detail. The system first attempts to determine whether a video is real or artificially generated. Subsequently, it can distinguish, among other things, whether a video was created via text-to-video or image-to-video, which model family was used, and which developer the model comes from. Finally, at the finest level, SAGA attempts to identify the specific generator. Movement as an AI fingerprintParticularly interesting about the method used by SAGA is that it examines not just individual frames, but sequences of images that contain additional information about how objects and image content change from frame to frame. Since different video generators leave characteristic patterns in these temporal sequences, according to the researchers, SAGA focuses precisely on this point. Using so-called Temporal Attention Signatures (T-Sigs), the researchers make visible the temporal patterns SAGA has learned from various generators. To achieve this, frame-to-frame attention values from multiple videos by the same generator are averaged. This creates a kind of temporal fingerprint that differs quite visibly between different AI models.  AI fingerprints of various video AI models via SAGA SAGA was tested on public datasets with a total of 19 different AI video generators. In the simple distinction between real and artificial videos, the method achieved an accuracy of 99.94 percent within the DeMamba dataset—an exceptionally high value, though achieved under controlled benchmark conditions. Also noteworthy is a test involving videos from as-yet-unknown AI models: When SAGA was trained only on videos from ten known models and then tested on clips from nine unknown ones, the accuracy was still 99.86 percent. When attributing a video to the specific video AI model used, SAGA achieved 94.99 percent. However, there are significant differences between individual models. The biggest drawback of the benchmarks listed in the paper: many current video generators such as Veo 3, Sora 2, Kling, or Runway Gen-4 are missing from the datasets used, which are primarily based on older models.
It also remains open how reliably such AI fingerprints survive multiple or relatively heavy video compression. Social networks in particular frequently re-compress uploaded videos and could thus blur the very subtle traces that SAGA relies on. Once SAGA is available as a practical tool, it will become clear how well it performs in realistic tests detecting videos from TikTok, YouTube, and Facebook. So far, SAGA is a research project and not a publicly available analysis tool. The academic paper is available for free via arXiv. Bild zur Newsmeldung:
 �more infos at bei arxiv.orgdeutsche Version dieser Seite: SAGA - Neues Tool erkennt KI-Videos mit 95% Genauigkeit |
 |
|
|