MiniWorld: Democratizing the Training of Video World Models from Scratch

MiniWorld presents a lightweight, transparent, and reproducible framework for training streaming video world models from scratch. The framework uses a block-causal Video Diffusion Transformer to autoregressively predict future observations from historical observations and control signals. The paper targets a gap in current practice: many strong systems adapt pretrained video generators through complex post-training or distillation pipelines, while bidirectional pretraining can mismatch causal, streaming inference. MiniWorld is intended as an end-to-end baseline that can be trained with modest computational resources, potentially making world-model research easier to reproduce and extend. The supplied abstract is truncated, so implementation details and evaluation results require verification in the full paper.