4D Driving Scene Generation With Stereo Forcing

arXiv 25.9 2025 56.3 method, application

TLDR

PhiGenesis generates temporally consistent 4D driving scenes from multi-view video using a two-stage framework with Stereo Forcing for geometric uncertainty.

Reasoning

The paper presents a novel unified framework for 4D scene generation that bridges video generation and novel view synthesis, with strong technical contributions in the Stereo Forcing conditioning strategy. However, the abstract lacks explicit mention of real-world datasets or benchmarks, and the evaluation details are incomplete, making it difficult to fully assess generalizability.

Read-first score

Read-first score 56.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.

Recency 8%

86.7

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

Methodology quality 25%

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

Topical relevance 42%

47.1

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%

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: 325.

Keyword Scores

generative world model

world dynamics prediction

world model

video world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Range-view adapter for feed-forward 4D reconstruction from multi-view images
Geometric-guided video diffusion model using rendered historical 4D scenes as priors
Stereo Forcing conditioning strategy to address geometric exposure bias and enhance temporal coherence

Methodology

PhiGenesis is a two-stage framework for 4D driving scene generation. In the first stage, a pre-trained video VAE with a novel range-view adapter performs feed-forward 4D reconstruction from multi-view image sequences, outputting temporally continuous 4D Gaussian splatting representations. In the second stage, a geometric-guided video diffusion model uses rendered historical 4D scenes as priors to generate future views conditioned on trajectories, with Stereo Forcing dynamically adjusting generative influence based on uncertainty-aware perturbations during denoising.

Key Results

The method achieves state-of-the-art performance in appearance and geometric reconstruction, temporal generation, and novel view synthesis (NVS), while also delivering competitive performance in downstream evaluations.

Resources

Paper Project