Tencent Hunyuan and Peking University release MovieRoll: trained on 15-second clips, it stays competitive at 30 seconds of streaming video
Synopsis
Streaming video generation must hold quality beyond the training horizon while handling self-generated history and growing cache costs. MovieRoll, from Tencent Hunyuan, Peking University and collaborators, chains streaming fine-tuning, long-rollout distillation and long-horizon inference into one base-to-distilled path: window-wise teacher forcing denoises neighboring chunks bidirectionally inside a window with causal access to history, and hierarchical KV-cache compression keeps different amounts of history at different depths. On HeliosBench, MovieRoll-Base remains competitive at 30 seconds after training on videos of up to 15 seconds, and MovieRoll-Distilled takes the highest total score at 10, 30 and 60 seconds for single events and at 30 and 60 seconds for multi-event generation.
Interpretation
Long-form streaming generation can now continue past the length of the training clips, and the release provides both a multi-step base model and a few-step distilled model. Among earlier multi-step streaming routes, Resampling Forcing evaluated durations only up to 15 seconds and Helios-Base used bidirectional continuation without KV caching, while the self-rollout distillation of Self Forcing and LongLive mainly targeted few-step generators. On HeliosBench, MovieRoll-Base remains competitive at 30 seconds after training on videos of up to 15 seconds, and MovieRoll-Distilled scores higher in total than the compared methods at 10, 30 and 60 seconds for single events and at 30 and 60 seconds for multi-event generation.
The base stage turns a pretrained bidirectional video model into a streaming model that generates over a rolling window, where neighboring chunks refine one another inside the window and only generated history is read outside it. Chunk-wise teacher forcing denoises each new chunk from pure noise, which encourages over-reliance on preceding content and lets historical errors accumulate; window-wise teacher forcing extends joint rolling-window denoising to base-model training and supervises several windows in parallel under their own event prompts and clean histories. After 10000 training steps on the same data, window-wise teacher forcing scores higher in total than chunk-wise teacher forcing and a chunk-wise variant given extra self-resampling training at both 10 and 30 seconds, with a larger lead at 30 seconds.
The history cache is retained hierarchically by network depth, keeping all history in shallow layers and only the sink and most recent chunk in deeper layers, which compresses memory without giving up generation quality. Full-history KV caching keeps raising memory and computation as video lengthens, while a uniform sliding window discards intermediate history at every layer. Relative to keeping full history at all layers, this allocation reduces KV-cache memory by up to 78% and peak denoising-step latency with CFG by approximately 53%, and it achieves the highest total score and drifting naturalness among the five tested allocations.
The few-step generator is trained on long self-rollouts, and each parameter update aggregates supervision across temporal positions of a whole rollout rather than a single supervision window. LongLive-style training updates from individual supervision windows, whereas CuLT-DMD holds parameters fixed throughout a rollout and averages local gradients before updating, retaining only one local computation graph per update. Starting from the same MovieRoll-Base checkpoint, CuLT-DMD scores higher in total than LongLive-style training and improves dynamics and drifting naturalness; extending Phase 2 rollouts from 241 to 721 frames further improves drifting aesthetics and drifting naturalness.
Perspective
The demonstrated scope is text-to-video streaming generation on HeliosBench: single events at 10, 30 and 60 seconds and multi-event generation at 30 and 60 seconds, with the base model trained on videos of up to 15 seconds and the distilled model trained on 241-frame and 721-frame rollouts. Training uses an internal video corpus that cannot be released because of confidentiality restrictions, and evaluation prompts are rewritten by its Streaming Video Prompt Rewriting Skill.
MovieRoll-Distilled has not yet achieved real-time generation: even with SageAttention and eight H20 GPUs, a 60-second video takes 152.2 seconds, or 9.47 FPS, against the 24 FPS needed for real-time playback. MovieRoll-Base is insufficiently stable at 60 seconds, so only the distilled model is evaluated at that duration, and the distilled model receives lower semantic and drifting-naturalness ratings at the longest evaluated duration than at the shortest.
