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

MiniWorld: Democratizing the Training of Video World Models from Scratch

arXiv 2026 67.5 method

TLDR

MiniWorld is a lightweight, reproducible framework to train autoregressive video world models from scratch using block-causal Video Diffusion Transformers.

Reasoning

The paper's strength lies in addressing the complexity and resource demands of prior methods by offering a fully reproducible baseline trained from scratch. However, the abstract lacks explicit mention of real-world benchmarks or empirical evaluations, and the claims about democratization are not supported by concrete results in the visible text.

Read-first score

Read-first score 67.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 59.

Recency 8%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 42%
84.3

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

Methodology quality 25%
50

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

Reproducibility 25%
46

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 289.

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
6

Deep Analysis

Innovations

  • A lightweight, reproducible framework for training streaming video world models from scratch without relying on pretrained video generation models.
  • Block-causal Video Diffusion Transformer with Flow Matching in the latent space of a pretrained Video VAE.
  • Chunk-wise non-decreasing noise schedule and two-stage continued training (based on Diffusion Forcing) for improved temporal modeling and stability.
  • Rolling KV cache with pipelined asynchronous denoising for efficient streaming generation under bounded computation.
  • Full training achievable within several days on a single 8-GPU server, enabling accessibility.

Methodology

MiniWorld uses a block-causal Video Diffusion Transformer trained with Flow Matching in the latent space of a pretrained Video VAE, incorporating a chunk-wise non-decreasing noise schedule and two-stage continued training for stability. Inference employs a rolling KV cache and pipelined asynchronous denoising for efficient autoregressive streaming generation.

Key Results

The entire MiniWorld model can be trained from scratch within several days on a single 8-GPU server, demonstrating accessibility and reproducibility.

Resources

Paper

Tags