StreamEval, introduced in a new arXiv preprint, is a unified benchmark for evaluating interactive streaming video generation models. The authors note that as streaming video generation advances, real-time interaction with generated video is becoming increasingly important. StreamEval aims to provide a common evaluation framework for this emerging capability.
The abstract describes the benchmark as spanning five areas, though the preprint's summary does not detail what those areas are. The paper is listed under arXiv identifier 2610.28402v1.
Because the source is only an abstract, the specific tasks, metrics, and baselines included in StreamEval are not yet available from this description.