Submitted to arXiv on August 24, 2026 (paper 2608.23383) by Nan Duan, Haoyang Huang, Weiyang Jin, Haoran Li, Yaowei Li, Yuming Li, Yijun Liu, Xin Lu, Xiaoxiao Ma, Yanwen Ma, Yaofeng Su, Yilang Sun, Haoyu Wang, Zeyue Xue, Songchun Zhang and Junhao Zhuang. The framing is that video generation is moving past isolated clips toward long-form narratives and interactive worlds, which changes the hard problem: a model now has to preserve identities, obey user controls, and stay stable over extended rollouts rather than merely produce one good shot.
The system, JoyAI-Echo-1.5, is presented as a unified audio-visual generator with two purpose-built variants. The long-video variant introduces composable cross-shot memory that aggregates visual evidence from multiple prior shots together with speaker cues derived from speech-filtered full-shot audio. That combination is what lets character appearance and voice identity persist across flexible mixtures of text, image and memory conditioning - the model can be told who someone is by showing it, describing it, or pointing back at an earlier shot.
The world-model variant addresses interaction rather than narrative. It converts heterogeneous navigation inputs into calibrated metric 6-DoF camera trajectories and injects them through a geometry-aware conditioning pathway, so the same model responds to different control schemes without being retrained per controller. To make long-horizon generation affordable, the authors convert a bidirectional audio-visual backbone into a causal few-step generator using progressive teacher forcing plus short- and long-horizon Self-Gradient Forcing applied to the model’s own rollouts.
The abstract reports strong performance in both settings without publishing headline metrics, so this belongs in the file as an architectural direction rather than a scoreboard entry. The direction is the point: cross-shot memory for identity and calibrated camera trajectories for control are the two capabilities that separate a clip generator from something that could carry a scene, a game world, or a synthetic character across an entire session.