AI 视频能拍多长,不再由训练时长决定
视频生成模型有个隐形天花板:训练时喂的片段多长,它大概就只能生成多长,再往后画面开始崩。这篇的做法是让模型在生成过程中不断「回看自己刚拍过的内容」,并把这段历史压缩成更省内存的形式——浅层保留全部、深层只留开头和最近几帧,缓存直接砍掉近八成,画质反而更好。训练上他们让模型自己滚出长序列、再按时间位置统一打分更新,一次只保留一小段计算图,于是几步就能出片的快速模型也能被拉到长程训练。结果是在最长只见过 15 秒素材的前提下,基础版撑到 30 秒仍有竞争力,蒸馏版在 60 秒内的单事件、多事件评测里总分最高,画面自然度和动作活跃度都保住了。做短视频、广告预演、需要连续长镜头的人,这是那种「天花板被抬高」的进展,不是明天就能上手的功能。
📄 原文摘要(英文)
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 establish a robust multi-step base through Window-wise Teacher Forcing (WinTF), which jointly denoises chunks at progressive noise levels with bidirectional attention within each window and causal access to preceding history. To retain historical context efficiently, hierarchical KV-cache compression (HieraKV) exploits layer-wise attention patterns, retaining full history in shallow layers and sink--recent history in deeper layers. This reduces cache memory by up to 78% while improving generation quality. Building on this base, cumulative long tuning with distribution matching distillation (CuLT-DMD) holds parameters fixed throughout each long self-rollout and averages local gradients before updating. Each update therefore combines supervision across temporal positions while retaining only one local computation graph, enabling efficient long-rollout training of a few-step generator. To extend generation beyond the training horizon, KV recomputation bounds context for both models, while guidance annealing (AnnealCFG) reduces guidance toward the clean endpoint to stabilize MovieRoll-Base. Complementing these modeling advances, FBCA attention and PS-VAE decoding support efficient training and streaming output. On HeliosBench, MovieRoll-Base remains competitive at 30s after training on videos of up to 15s, and MovieRoll-Distilled achieves the highest total scores among the compared methods across single- and multi-event evaluations up to 60s, combining sustained naturalness with active motion.