⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

Learning to Generate 4D LiDAR Sequences

ICCVW 25 2025 48.3 method

TLDR

LiDARCrafter generates editable 4D LiDAR sequences from language using scene graphs and diffusion models, achieving state-of-the-art fidelity and temporal consistency.

Reasoning

The paper presents a novel framework for LiDAR sequence generation with explicit controllability and temporal coherence, supported by a new benchmark. However, it is limited to LiDAR modality and lacks interactive simulation or broader world model capabilities.

Read-first score

Read-first score 48.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 27.

Recency 8%
86.7

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

Methodology quality 25%
70

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

Topical relevance 42%
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 25%
30

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: 395.

Keyword Scores

generative world model
8
world model
7
world dynamics prediction
5
world simulator
3
video world model
2
interactive world model
1
model-based reinforcement learning world model
1

Deep Analysis

Innovations

  • Unified framework converting free-form language into editable LiDAR sequences
  • Parsing instructions into ego-centric scene graphs
  • Tri-branch diffusion model for object layouts, trajectories, and shapes
  • Range-image diffusion model for initial scan generation
  • Autoregressive module for temporally coherent sequence extension
  • Object-level editing capabilities (insertion, relocation)
  • EvalSuite benchmark with scene-, object-, and sequence-level metrics

Methodology

The methodology parses free-form language instructions into ego-centric scene graphs. A tri-branch diffusion model transforms these graphs into object layouts, trajectories, and shapes. A range-image diffusion model generates the initial LiDAR scan, and an autoregressive module extends it into a temporally coherent sequence. Evaluation is performed using the proposed EvalSuite benchmark with multi-level metrics.

Key Results

On nuScenes, LiDARCrafter achieves state-of-the-art fidelity, controllability, and temporal consistency.

Resources

Paper

Tags