LaMo: Self-Supervised Latent Motion Priors for Physical Realism in Video Generation

arXiv 2026 51.6 method

TLDR

LaMo extracts self-supervised motion priors from unlabeled videos to improve physical realism in video diffusion models, achieving state-of-the-art on physics benchmarks.

Reasoning

The paper presents a novel self-supervised method that leverages unlabeled video data to enhance motion consistency in video generation, with strong empirical results on VideoPhy and VBench. However, it focuses narrowly on motion cues and does not address broader physical simulation or interactive capabilities.

Read-first score

Read-first score 51.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 27.

Recency 6%

100

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

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.84977172

Methodology quality 18%

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

Topical relevance 29%

38.6

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 18%

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 299.

Keyword Scores

world simulator

world dynamics prediction

video world model

world model

generative world model

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Self-supervised extraction of motion cues from unlabeled videos already used to train video diffusion models, without external simulators or curated physics data.
Formulation of a latent motion prior over frame-to-frame latent changes, conditioned on the current latent and prompt.
Two lightweight readouts: a macro motion drift used as a Motion Drift Loss during training, and a learned micro motion field used as Motion Prior Guidance during sampling.
Plug-and-play integration with existing video diffusion backbones, requiring no architectural or I/O changes.

Methodology

LaMo extracts a latent motion prior from unlabeled videos by modeling frame-to-frame latent changes conditioned on the current latent and prompt. It employs two readouts: a macro motion drift loss for training and a micro motion field for sampling guidance, both designed to be compatible with any video diffusion backbone without architectural modifications. The method is evaluated on VideoPhy, VideoPhy2, and VBench benchmarks using CogVideoX backbones, comparing against recent physics-aware baselines that use external supervision.

Key Results

On VideoPhy and VideoPhy2, LaMo improves CogVideoX backbones and outperforms recent physics-aware baselines that rely on external supervision. On VBench, it preserves overall generation quality while improving motion-related dimensions.

Limitations

The self-supervised approach may not capture all physical phenomena, as it relies solely on motion cues from unlabeled video data.
The method is only evaluated on CogVideoX backbones; generalization to other video diffusion architectures is not demonstrated.
While improving motion consistency, the paper does not claim full physical realism, and limitations in physical fidelity may persist.

Resources

Paper