Paper 2512.13604
LongVie 2: Multimodal Controllable Ultra-Long Video World Model
- Published
- Dec 2025
- Research lab
- Independent
- Citations
- 6
- GitHub
- 335 stars
01 In brief
Summary
LongVie 2 is an end-to-end autoregressive framework for controllable ultra-long video generation, extending pretrained diffusion backbones (Wan2.1-I2V-14B) into a video world model.
It is trained in three progressive stages: (1) multi-modal guidance integrating dense (depth maps) and sparse (point maps) control signals via a ControlNet-style architecture with degradation-based balancing; (2) degradation-aware training that simulates VAE and diffusion-induced corruption on input frames to bridge the training-inference gap; and (3) history-context guidance using preceding frames as conditional context, with temporal regularization losses, to maintain cross-clip consistency.
Training-free strategies (unified noise initialization and global normalization) further improve inter-clip coherence.
The authors introduce LongVGenBench, a benchmark of 100 one-minute 1080p videos covering real-world and synthetic scenes.
Experiments show LongVie 2 achieves state-of-the-art performance in controllability, temporal consistency, and visual fidelity, supporting generation up to 3–5 minutes.
Ablations confirm each stage contributes to quality, controllability, and consistency.
The work addresses limitations of current world models in controllability and temporal scalability, marking progress toward unified video world modeling.
02 From the paper
Abstract
Building video world models upon pretrained video generation systems represents an important yet challenging step toward general spatiotemporal intelligence. A world model should possess three essential properties: controllability, long-term visual quality, and temporal consistency. To this end, we take a progressive approach-first enhancing controllability and then extending toward long-term, high-quality generation. We present LongVie 2, an end-to-end autoregressive framework trained in three stages: (1) Multi-modal guidance, which integrates dense and sparse control signals to provide implicit world-level supervision and improve controllability; (2) Degradation-aware training on the input frame, bridging the gap between training and long-term inference to maintain high visual quality; and (3) History-context guidance, which aligns contextual information across adjacent clips to ensure temporal consistency. We further introduce LongVGenBench, a comprehensive benchmark comprising 100 high-resolution one-minute videos covering diverse real-world and synthetic environments. Extensive experiments demonstrate that LongVie 2 achieves state-of-the-art performance in long-range controllability, temporal coherence, and visual fidelity, and supports continuous video generation lasting up to five minutes, marking a significant step toward unified video world modeling.