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

DiST-4D: Disentangled Spatiotemporal Diffusion with Metric Depth for 4D Driving Scene Generation

ICCV 25 2025 58.6 method, application

TLDR

A disentangled spatiotemporal diffusion framework using metric depth for 4D driving scene generation, achieving SOTA in temporal prediction and novel view synthesis.

Reasoning

The paper introduces a novel decomposition of 4D generation into two diffusion processes with cycle consistency, leveraging metric depth for geometric consistency. Strengths include strong empirical results on temporal and spatial tasks, but the abstract lacks explicit details on real-world datasets and limitations.

Read-first score

Read-first score 58.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

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=evaluation,experiment,metric

Topical relevance 42%
58.6

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%
38

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 201.

Keyword Scores

world dynamics prediction
9
generative world model
8
video world model
8
world model
7
world simulator
6
interactive world model
2
model-based reinforcement learning world model
1

Deep Analysis

Innovations

  • First disentangled spatiotemporal diffusion framework for 4D driving scene generation
  • Leverages metric depth as the core geometric representation to connect temporal and spatial synthesis
  • Cycle consistency mechanism with forward-backward rendering constraint to reduce generalization gap between observed and unseen viewpoints

Methodology

DiST-4D decomposes 4D driving scene generation into two diffusion processes: DiST-T predicts future metric depth and multi-view RGB sequences from past observations, and DiST-S enables spatial novel view synthesis by training only on existing viewpoints while enforcing cycle consistency via a forward-backward rendering constraint. Metric depth provides a view-consistent geometric representation that generalizes to unseen perspectives.

Key Results

DiST-4D achieves state-of-the-art performance in both temporal prediction and novel view synthesis tasks, and delivers competitive performance in planning-related evaluations.

Tags