Video Generation with Predictive Latents

arXiv 2026 41.3 method

TLDR

Introduces PV-VAE, a video VAE with predictive reconstruction that improves generation quality and convergence speed.

Reasoning

Strengths: Simple yet effective predictive objective that enhances video dynamics understanding, with clear empirical gains (52% faster convergence, 34.42 FVD improvement). Weaknesses: Limited to UCF101 benchmark; no discussion of failure cases or broader real-world deployment.

Read-first score

Read-first score 41.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 22.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

70.5

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.70458759

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Topical relevance 29%

31.4

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

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierBridge

Rank sensitivity

Stability: volatile; rank range: 243.

Keyword Scores

world dynamics prediction

video world model

world model

generative world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Predictive reconstruction objective that unifies predictive learning with video reconstruction
Randomly discarding future frames and encoding only partial past observations while training decoder to reconstruct observed and predict future frames
PV-VAE model that improves latent space temporal coherence and motion priors

Methodology

The method introduces a predictive reconstruction objective for video VAEs. During training, future frames are randomly discarded, and only partial past observations are encoded. The decoder is trained to simultaneously reconstruct the observed frames and predict the future ones, encouraging the latent space to encode temporally predictive structures.

Key Results

PV-VAE achieves 52% faster convergence and a 34.42 FVD improvement over Wan2.2 VAE on UCF101. It also shows favorable scalability and consistent gains in downstream video understanding.

Resources

Paper