Skip to content

Source paper

MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos

Yunyang Ge, Weijie Kong, Ryan Xu and 9 more

Yunyang Ge, Weijie Kong, Ryan Xu, Yue Wu, Qi Tian, Jiangfeng Xiong, Bin Lin, Mingzhe Zheng, Shenghai Yuan, Zheyuan Liu, Jianwei Zhang, Li Yuan

arXiv 2611.21339 · 202613 upvotes on Hugging Face

Research ideas from this paper (5)

Generated by AI. The paper’s authors did not write these.

Quantized Attention, Denoising-Decoder Pipelining, and Sharded VAE Bottlenecks for Real-Time Streaming Video

Accelerate streaming video generation toward real-time playback by quantizing transformer weights, sharding decoder bottleneck attention across GPUs, and overlapping denoising with decoding.

MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos
🟠Medium💡Moderate InnovationMultiple GPUsRelated work found#Efficiency#Generative media

Closed-Loop Visual-Feedback Agent for Complex Camera Trajectories and Multi-Event Continuity

Build a visual-feedback agent that inspects newly generated video frames to dynamically update event prompts and plan complex camera trajectories without losing scene continuity.

MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos
🟠Medium🔬Significant ResearchMultiple GPUsNo close match found#Agents#Multimodal

Duration-Conditioned and Shot-Triggered Rolling-Window Pretraining with Finer Chunk Granularity

Condition rolling-window video models on total duration, shot-boundary triggers, and finer chunk steps so long-form generation avoids short-clip pacing bias and handles multi-shot transitions.

MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos
🔴Hard🔬Significant ResearchLarge clusterRelated work found#Generative media#Training

Prompt-Relevant Historical Memory Retrieval and Entropy-Calibrated Layer-Head KV Compression

Replace fixed-rule cache eviction with prompt- and visual-relevance retrieval and entropy-calibrated per-head compression to reduce semantic drift over minute-long video rollouts.

MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos
🟡Fairly easy💡Moderate InnovationMultiple GPUsRelated work found#Generative media#Retrieval & memory

Stratified Minute-Scale Rollout Distillation and Self-Resampled Base Stabilization for 60s+ Generation

Extend self-rollout distillation to minute-scale multi-event horizons via stratified late-window supervision and scheduled-sampling base tuning so both base and fast generators remain stable.

MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos
🔴Hard💡Moderate InnovationLarge clusterRelated work found#Generative media#Training

AI summary

MovieRoll introduces a base-to-distilled framework for streaming long-form video generation that sustains visual quality, active motion, and prompt continuity up to 60 seconds while bounding historical KV-cache memory. By combining parallel rolling-window fine-tuning, hierarchical layer-wise cache compression, cumulative long-rollout distribution matching distillation, and streaming decoding infrastructure, it outperforms existing streaming baselines on single-event and multi-event HeliosBench evaluations.

How it works (AI summary)

Starting from an 8.3B-parameter bidirectional MMDiT backbone (HunyuanVideo-1.5-480P-T2V), the authors first train a multi-step streaming base model (MovieRoll-Base) using Window-wise Teacher Forcing (WinTF), where chunks within an active window attend bidirectionally at progressive noise levels while causally attending to clean history, accelerated by Flash Block-wise Causal Attention (FBCA). They then initialize a 5-step student from annealed-CFG ODE trajectories and train MovieRoll-Distilled via Cumulative Long Tuning with DMD (CuLT-DMD), keeping weights fixed across 241- to 721-frame self-rollouts and averaging local gradients before each update. At inference, Hierarchical KV-cache compression (HieraKV) keeps full history only in the first 6 blocks and sink-plus-recent chunks in the remaining 48 blocks, threshold-triggered KV recomputation rebuilds the cache at 18 chunks or prompt switches, AnnealCFG decays guidance from 4 to 1 for the base model, and Parallel Streaming VAE (PS-VAE) decodes chunks incrementally across GPUs.