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

Dreamland: Controllable World Creation with Simulator and Generative Models

arXiv 25.6 2025 58.2 method, system

TLDR

Dreamland combines physics simulator and generative models for controllable, photorealistic world creation, improving image quality and controllability.

Reasoning

Strengths include a novel hybrid framework with layered abstraction and strong quantitative improvements. Weaknesses are limited detail on embodied agent training results and reliance on pretrained models.

Read-first score

Read-first score 58.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 31.

Recency 8%
86.7

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

Methodology quality 25%
80

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

Reproducibility 25%
50

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

Topical relevance 42%
44.3

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 anchor

Rank sensitivity

Stability: volatile; rank range: 374.

Keyword Scores

world simulator
8
generative world model
6
video world model
5
interactive world model
4
world model
3
model-based reinforcement learning world model
3
world dynamics prediction
2

Deep Analysis

Innovations

  • Hybrid world generation framework combining physics-based simulator and large-scale pretrained generative models for element-wise controllability
  • Layered world abstraction encoding pixel-level and object-level semantics and geometry as intermediate representation to bridge simulator and generative model
  • D3Sim dataset for training and evaluation of hybrid generation pipelines

Methodology

Dreamland uses a layered world abstraction that encodes both pixel-level and object-level semantics and geometry as an intermediate representation to bridge a physics-based simulator and a large-scale pretrained generative model. This approach enables granular control from the simulator while leveraging the photorealistic output of the generative model. The framework is designed for off-the-shelf use of existing and future pretrained models, and is trained and evaluated on the constructed D3Sim dataset.

Key Results

Dreamland outperforms existing baselines with 50.8% improved image quality and 17.9% stronger controllability, and demonstrates potential to enhance embodied agent training.

Resources

PaperProject

Tags