PKU & Tencent Hunyuan's MovieRoll extends streaming video generation to 60s with 78% KV-cache savings
Yunyang Ge · hf · 2026-11-17
Researchers from Peking University, Tencent Hunyuan, and HKUST present MovieRoll, a base-to-distilled framework for streaming long-video generation.
- WinTF: window-wise teacher forcing with bidirectional in-window attention and causal history access builds a robust multi-step base.
- HieraKV: hierarchical KV-cache compression keeps full history in shallow layers and sink+recent history in deeper layers, cutting cache memory up to 78% while improving quality.
- CuLT-DMD freezes parameters during each long self-rollout and averages local gradients, enabling efficient long-rollout distillation of a few-step generator.
- KV recomputation plus AnnealCFG guidance annealing extends generation beyond the training horizon; FBCA attention and PS-VAE support streaming output.
On HeliosBench, MovieRoll-Base stays competitive at 30s after training on ≤15s videos, and the distilled model tops comparisons up to 60s.
More from Multimodal
- Ink Studio launches free on GitHub for 3D pixel-art Gaussian splat animation — ring_hyacinth · 2026-11-17
- UMM-JEPA: predicting continuous perceptual representations beats discrete codes for unified multimodal models — Yucheng Zhou · 2026-11-17
- GlyphAtlas disentangles stroke detail from layout for accurate scene text synthesis — Luozheng Qin · 2026-11-17
- 'veo-experimental' spotted in hidden docs, suggesting a new Google Veo model is in the works — koltregaskes · 2026-11-17
- Built with ChatGPT: Forge extension tames 30,000-LoRA collections with incremental loading — Artistic-Account-423 · 2026-11-17
- AI-generated Godzilla-style kaiju video of Tokyo street panic impresses — Careful-Loquat-8117 · 2026-11-17