Video world models are a class of AI systems that predict future video frames based on past observations and control signals, enabling long-horizon generation through autoregressive state transitions. Unlike conventional video generation models that focus on visual appearance and motion, world models learn the underlying dynamics of an environment under agent actions, making them a foundation for embodied AI and interactive simulation.
Recent progress in this area has largely depended on adapting pretrained video generation models through post-training or distillation. While effective, these approaches often require complex training pipelines, substantial computational resources, and suffer from a mismatch between bidirectional pretraining and causal streaming inference. Although some studies have shown that training autoregressive video world models from scratch is feasible and scalable, the community still lacks a lightweight, transparent, and fully reproducible baseline that can be trained end-to-end with modest resources.
To address this gap, researchers at Hugging Face present MiniWorld, a reproducible framework for training streaming video world models from scratch. MiniWorld employs a block-causal Video Diffusion Transformer trained with Flow Matching in the latent space of a pretrained Video VAE. Building on Diffusion Forcing, it adopts a chunk-wise non-decreasing noise schedule and two-stage continued training to improve temporal modeling and stability.
During inference, MiniWorld combines a rolling KV cache with pipelined asynchronous denoising for efficient streaming generation under bounded computation. The entire model can be trained within several days on a single 8-GPU server. By releasing the training and inference codebase and pretrained checkpoints, the team hopes MiniWorld will facilitate future research on video world modeling.