MiniWorld: Democratizing the Training of Video World Models from Scratch

Video world models predict future observations conditioned on historical observations and control signals. Unlike conventional video generators, which mainly capture visual appearance and motion, they learn the dynamics governing how an environment evolves under agent actions — a fundamental building block for embodied AI and interactive simulation.

Recent progress has largely relied on adapting pretrained video generators through post-training or distillation. These pipelines are complex, costly, and inherit a mismatch between bidirectional pretraining and causal streaming inference. Training from scratch is feasible and scalable, yet the community still lacks a lightweight, transparent, reproducible baseline.

MiniWorld fills that gap. It uses a block-causal Video Diffusion Transformer trained with Rectified Flow in the latent space of a pretrained video VAE, and builds on Diffusion Forcing with a chunk-wise non-decreasing noise schedule and two-stage continued training. At inference, a rolling KV cache and pipelined asynchronous denoising keep streaming generation within a bounded compute budget.

The entire model trains in several days on a single 8-GPU server. We release the complete training and inference codebase, pretrained checkpoints, and implementation details as an accessible baseline for research on long-horizon generation, temporal memory, and train-test alignment.