以下为上游提供的源文稿,可能存在识别错误。
乔木转写是基于源文稿的中文提炼,不是逐字翻译。请核对原文中的人名、数字与观点。
选一篇,慢慢读
从左侧挑一篇,正文会在这里排成书页。
Streaming video generation must sustain quality beyond the training horizon while handling self-generated history and growing cache costs. To address these challenges, we present MovieRoll, a base-to-distilled framework that connects streaming fine-tuning, long-rollout distillation, and long-horizon inference. We first
阅读原文 · Yunyang Ge, Weijie Kong, Ryan Xu, Yue Wu, Qi Tian, Jiangfeng Xiong, Bin Lin, Mingzhe Zheng, Shenghai Yuan, Zheyuan Liu, Jianwei Zhang, Li Yuan
向阳乔木 · opencode-go/deepseek-v4.1-flash
长视频生成有个尴尬的现状:模型训练时看的是十几秒片段,真正上线却要连续吐几十秒甚至更久。一旦超出训练时长,画面开始漂、动作变僵、显存被历史帧吃掉,质量是断崖式下滑的。 MovieRoll 想处理的就是这段“训练地平线之外”的空白。它把流式微调、长 rollout 蒸馏、长时推理三件事串成一个从 base 到 distilled 的框架, 核心主张是:长视频能...