NVIDIA Research 发布 WorldTrace:让交互式视频世界模型记住去过的场景
做交互式视频或世界模型的开发者可以留意,NVIDIA Research 这套 WorldTrace 不用重训练就能让模型记住回访过的场景,在 LoopBench 上比滑动窗口方法提升 32% 召回率。如果你在搭 Genie 类交互生成管线,这份 NeurIPS 论文的思路可以直接借鉴。
NVIDIA Research 推出 WorldTrace,一种无需重训练就能让交互式视频世界模型记住已探索场景的方法,解决用户离开再返回同一场景时画面不一致的问题。
在场景回溯基准 LoopBench 上,相比标准滑动窗口方法,重访场景的召回率提升 32%。其核心机制是在模型探索过程中追踪并保存场景记忆信息,使空间状态在多次进入时保持连贯。
该方法以即插即用方式工作,不改底层模型参数。论文已被 NeurIPS 2026 收录,对做世界模型和交互式视频生成的团队有直接参考价值。
Look away and return to the same scene. For interactive video world models, preserving that consistency is a challenge. WorldTrace from @NVIDIA Research helps models remember places without retraining, keeping spaces more consistent as users explore. On LoopBench, a benchmark for scene recall, it improves recall of revisited scenes by 32% over the standard sliding-window approach. Read our #NeurIPS2026 paper ➡️ nvda.ws/4knUqaA
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include WorldTrace, NVIDIA, world models, video generation.
Source attribution: curated from NVIDIA AI (@NVIDIAAI).