Evoke uses bounded scene memory for long-video sessions
In plain words
Evoke is a video world model designed to keep generating as a camera moves and instructions change. It creates each new video segment using recent frames and a limited store of scene geometry from earlier views. A separate teacher trains it on long sequences with changing instructions.
In tests, revisited views improved when the stored geometry covered the time spent away. The memory is limited, and the study found a specific brightness-stability gain from long-horizon teacher training, not better results on every measure.
The research
When an interactive video session runs for a long time, how can its generator remember a place the camera revisits without carrying an ever-growing history?
It must also respond to new instructions, even though a fast generator trained on short sequences may not learn how to handle changes across distant moments.
Evoke pairs a bounded external geometry bank with long-horizon teacher training to support three-step video generation over long sessions.
The authors compared matched three-step students trained from short- and long-horizon teachers. Across extended rollouts, the long-horizon student settled at 101% of its opening brightness, compared with 74% for the short-horizon student; seven of eight clips improved.
Because their content descriptor did not significantly differ, the comparison supports better resistance to brightness drift, not a general improvement in long-range consistency.
These tests support bounded recall and specific forms of stability, not permanent memory or consistent preservation of every scene detail. The bank retains only a finite window of geometry, and the teacher comparison did not show a broad improvement across all measures of long-range consistency.
Analysis
Deep
- Authors
- Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang
- License
- arXiv non-exclusive