Yume: An Interactive World Generation Model

arXiv 25.7 2025 75.7 method, system

TLDR

Yume generates interactive, dynamic worlds from images using keyboard control, with a framework including camera motion quantization, MVDT, and advanced sampling.

Reasoning

The paper presents a novel framework for interactive world generation with clear technical contributions (MVDT, AAM, TTS-SDE) and uses a dedicated dataset. However, the abstract lacks explicit quantitative results or comparisons, and the claim of 'remarkable results' is vague without empirical evidence.

Read-first score

Read-first score 75.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 50.

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=artifact,code,dataset,github

Topical relevance 42%

71.4

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

Methodology quality 25%

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

Field roles

FrontierMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 64.

Keyword Scores

interactive world model

generative world model

video world model

world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Camera motion quantization for stable training and user-friendly keyboard-based interaction
Masked Video Diffusion Transformer (MVDT) with a memory module for autoregressive infinite video generation
Training-free Anti-Artifact Mechanism (AAM) and Time Travel Sampling based on Stochastic Differential Equations (TTS-SDE) for improved visual quality and control
Synergistic optimization of adversarial distillation and caching mechanisms for model acceleration

Methodology

The framework comprises four components: camera motion quantization to discretize motions for stable training and keyboard control; a Masked Video Diffusion Transformer (MVDT) with a memory module enabling autoregressive infinite video generation; a sampler incorporating training-free Anti-Artifact Mechanism (AAM) and Time Travel Sampling based on Stochastic Differential Equations (TTS-SDE) for better quality and control; and model acceleration via adversarial distillation combined with caching. The model is trained on the high-quality Sekai world exploration dataset and evaluated on diverse scenes.

Key Results

The preview version of Yume generates a dynamic, interactive world from a single input image, allowing keyboard-based exploration, and achieves remarkable results across diverse scenes and applications.

Limitations

Preview version only supports image input, not text or video as originally aimed
Exploration is limited to keyboard actions, not peripheral devices or neural signals
The model is still under development with monthly updates planned, indicating incomplete functionality

Resources

Paper Code