DynamicCity: Large-Scale LiDAR Generation from Dynamic Scenes

ICLR 25 2025 52.5 method

TLDR

DynamicCity generates large-scale dynamic 4D LiDAR occupancy scenes using VAE and diffusion models with novel compression and generation techniques.

Reasoning

The paper introduces a novel 4D occupancy generation framework with efficient HexPlane representation and diffusion-based generation, showing strong quantitative gains. However, it focuses narrowly on LiDAR occupancy and does not address broader world model concepts like interaction, reinforcement learning, or video generation.

Read-first score

Read-first score 52.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 11.

Recency 8%

86.7

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

Reproducibility 25%

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

Methodology quality 25%

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

Topical relevance 42%

15.7

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

FrontierMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 686.

Keyword Scores

world dynamics prediction

generative world model

world model

world simulator

video world model

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Projection Module for compressing 4D features into six 2D feature maps for HexPlane construction, improving fitting quality by up to 12.56 mIoU
Expansion Squeeze Strategy to reconstruct 3D feature volumes in parallel, achieving up to 7.05 mIoU gain, 2.06x training speedup, and 70.84% memory reduction
Padded Rollout Operation to reorganize HexPlane's six feature planes into a squared 2D map for DiT-based diffusion generation
Versatile conditional generation capabilities including trajectory-driven, command-driven, inpainting, and layout-conditioned generation

Methodology

DynamicCity employs a VAE to learn a compact 4D representation called HexPlane, using a Projection Module to compress 4D features into six 2D maps and an Expansion Squeeze Strategy for efficient parallel 3D volume reconstruction. A DiT-based diffusion model then generates HexPlane via a Padded Rollout Operation that reorganizes the six planes into a squared 2D feature map, enabling various conditional generation tasks during diffusion or sampling.

Key Results

DynamicCity significantly outperforms existing state-of-the-art 4D occupancy generation methods on the CarlaSC and Waymo datasets, with the Projection Module yielding up to 12.56 mIoU gain and the Expansion Squeeze Strategy providing up to 7.05 mIoU gain, 2.06x training speedup, and 70.84% memory reduction.

Resources

Paper Code