This is a Plain English Papers summary of a research paper called StreamEval Tests Interactive Video Models Across 300 Timed Sessions. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
StreamEval tests whether streaming video systems follow requests over time
StreamEval evaluates 38 models across five interactive video-generation tasks, using 300 scripted sessions that last 15 to 300 seconds. Its central test is whether a system can act on timed user requests while preserving relevant scene or identity state, rather than producing only plausible frames. Ten metrics cover execution, consistency, visual fidelity, and audiovisual coordination.
The benchmark spans digital-human dialogue, avatar performance, world generation, instruction editing, and reference editing. Each case pairs an initial portrait, scene, or source video with timed inputs and explicit expectations about what should change and what must remain intact. For example, an editing session can add an object change, update another target, then remove an edit to restore the advancing source.
The overview groups five tasks across avatar interaction, world generation, and real-time editing. It makes clear that StreamEval evaluates varied interactive settings under one benchmark....
