UC Riverside highlighted SAGA on July 24, 2026, months after the paper first appeared in November 2025 and following its publication at CVPR 2026. The framework classifies synthetic videos at five levels, from real-versus-fake to a specific generator, and reports 94.99% generator-level accuracy with 0.5% labeled source data on its benchmark. These are research-dataset results, not proof of reliability on arbitrary real-world video.
UC Riverside highlighted SAGA, a research framework for attributing synthetic videos to their source generator, in a July 24 announcement. The timing matters: the paper first appeared on arXiv in November 2025 and was published at CVPR 2026, so the university article was a new explanation of existing research rather than the framework's first release.
SAGA, short for Source Attribution of Generative AI Videos, goes beyond a binary real-or-fake decision. Its authors evaluate attribution at five levels: authenticity, text-to-video versus image-to-video task, Stable Diffusion backbone version, development team and the specific generator.
How SAGA performs attribution
The framework combines spatial features from individual frames with temporal information across a video. Its Temporal Attention Signatures summarize frame-to-frame attention patterns that the authors use to visualize differences among generator classes.
Training happens in two stages. The researchers first train a real-versus-fake classifier using a larger pool of binary labels. They then adapt that representation to source attribution with hard-negative mining while using 0.5% of the source-labeled data per class.
The paper evaluates 19 video generators across the DeMamba and DVF public datasets. In its most specific generator-level task, the two-stage method reports 94.99% overall accuracy with 0.5% source-labeled data, compared with up to 97.41% under the paper's full-data setting. The paper also reports in-domain and cross-domain experiments.
What the results do and do not show
Those figures are benchmark results under the authors' datasets and class definitions. They do not establish that SAGA can reliably identify every model behind an arbitrary video found online. Compression, editing, model updates, generators absent from the training taxonomy and deliberate evasion can all change the operational problem.
The paper's own analyses show why task definition matters. Coarser representations can separate real, text-to-video and image-to-video content without cleanly separating every generator. Fine-grained attribution requires the source classes and hard negatives used during adaptation.
For practitioners, SAGA is best read as evidence that generator attribution can be framed at multiple levels and learned with relatively little fine-grained labeled data after broader pretraining. Independent replication, open-set testing and evaluation on newly released or post-processed generators would be needed before treating it as a production forensic identifier.
Key Points #
- 1UC Riverside's July 24 article highlighted research first posted in November 2025 and published at CVPR 2026.
- 2SAGA attributes synthetic video at five levels, ranging from authenticity to a specific generator.
- 3The paper reports 94.99% generator-level benchmark accuracy with 0.5% source-labeled data, but the result does not establish reliability on arbitrary real-world video.
Scoring Rationale #
SAGA reports strong multi-level attribution results with limited fine-grained labels, but the evidence remains benchmark-based and needs independent, open-set and post-processing validation before operational use.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.