StreamEval benchmark measures interactive streaming video generation across five tasks

Streaming video generation is advancing toward systems that generate video incrementally and respond to user inputs in real time, enabling applications like conversational avatars, controllable virtual worlds, and live video editing. These applications require more than generating plausible frames; they must maintain alignment with evolving user requests and context over time.

Sustained interaction demands that models execute new requests while preserving relevant prior content and state. For example, a conversational avatar must maintain identity and dialogue history while responding appropriately; a generated world must update controlled objects without disrupting unaffected content; and a video editor must apply and update edits while tracking the advancing source video.

Evaluating such systems is difficult because it requires measuring both the correctness of request execution and the consistency of preserved state, alongside visual fidelity and audiovisual coordination. Existing benchmarks cover parts of these needs but do not unify avatar interaction, world generation, and real-time video editing with explicit execution and preservation criteria across multiple tasks and durations.