LongVie 2: Multimodal Controllable Ultra-Long Video World Model

arXiv 25.12 2025 57.8 method

TLDR

LongVie 2 is a three-stage autoregressive framework for controllable, long-term video world modeling with state-of-the-art performance.

Reasoning

The paper presents a clear methodology with three training stages and introduces a new benchmark, demonstrating strong empirical results. However, it lacks explicit comparison to model-based RL world models and does not address interactive or simulator aspects beyond controllability.

Read-first score

Read-first score 57.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,benchmark,experiment

Topical relevance 42%

61.4

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 269.

Keyword Scores

video world model

world model

generative world model

world dynamics prediction

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Multi-modal guidance integrating dense and sparse control signals for implicit world-level supervision
Degradation-aware training on the input frame to bridge the gap between training and long-term inference
History-context guidance aligning contextual information across adjacent clips for temporal consistency
LongVGenBench, a comprehensive benchmark of 100 high-resolution one-minute videos covering diverse environments

Methodology

LongVie 2 is an end-to-end autoregressive framework trained in three stages: (1) multi-modal guidance using dense and sparse control signals to improve controllability, (2) degradation-aware training on input frames to maintain visual quality during long-term inference, and (3) history-context guidance to align contextual information across adjacent clips for temporal consistency. The model is evaluated on the introduced LongVGenBench benchmark and compared against baselines for long-range controllability, temporal coherence, and visual fidelity.

Key Results

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.

Resources

Paper