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

MaskGWM: A Generalizable Driving World Model with Video Mask Reconstruction

CVPR 25 2025 85.8 method, application

TLDR

MaskGWM combines diffusion transformers with MAE-style mask reconstruction for generalizable driving world models, enabling long-horizon and multi-view video prediction.

Reasoning

The paper introduces a novel integration of generative diffusion loss and feature-level mask reconstruction to improve generalization and predictive duration in driving world models. Strengths include clear architectural innovations (DiT with spatial-temporal masking) and validation on standard benchmarks. Weaknesses are the narrow focus on driving scenarios and lack of evidence for real-world deployment beyond dataset evaluation.

Read-first score

Read-first score 85.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 56.

Methodology quality 25%
100

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

Recency 8%
86.7

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

Reproducibility 25%
81

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

Topical relevance 42%
80

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: stable; rank range: 0.

Keyword Scores

world model
10
video world model
10
generative world model
9
world dynamics prediction
9
interactive world model
8
world simulator
7
model-based reinforcement learning world model
3

Deep Analysis

Innovations

  • Combining generation loss with MAE-style feature-level context learning
  • Scalable Diffusion Transformer (DiT) structure trained with extra mask construction task
  • Diffusion-related mask tokens to handle fuzzy relations between mask reconstruction and generative diffusion
  • Spatial-temporal mask construction using row-wise mask for shifted self-attention instead of masked self-attention in MAE
  • Row-wise cross-view module aligned with mask design
  • Two variants: MaskGWM-long for long-horizon prediction and MaskGWM-mview for multi-view generation

Methodology

The model uses a Diffusion Transformer (DiT) architecture with an additional mask construction task. It introduces diffusion-related mask tokens and employs row-wise masks for shifted self-attention to extend mask construction to the spatial-temporal domain, along with a row-wise cross-view module. Training and evaluation are conducted on standard benchmarks including Nuscenes, OpenDV-2K, and Waymo datasets, comparing against state-of-the-art driving world models.

Key Results

The method achieves notable improvements over state-of-the-art driving world models on normal validation (Nuscenes), long-horizon rollout (OpenDV-2K), and zero-shot validation (Waymo) datasets.

Resources

PaperCode

Tags