SV4D: Dynamic 3D Content Generation with Multi-Frame and Multi-View Consistency

arXiv 2024 28.8 method

TLDR

Stable Video 4D introduces a unified latent video diffusion model to generate multi-view consistent novel view videos from monocular video, enabling efficient dynamic NeRF 4D generation.

Reasoning

The paper proposes a unified diffusion model for multi-frame and multi-view consistent dynamic 3D generation, with strong empirical results and user studies. However, the abstract relies on the Objaverse dataset and does not clarify whether real-world dynamic scenes are included, limiting evidence of generalizability.

Read-first score

Read-first score 28.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 1.

Recency 8%

75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Methodology quality 25%

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

Reproducibility 25%

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

Topical relevance 42%

1.4

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

Field roles

Candidate

Rank sensitivity

Stability: volatile; rank range: 80.

Keyword Scores

video world model

world model

world simulator

generative world model

interactive world model

world dynamics prediction

model-based reinforcement learning world model

Resources

Paper