Peking University Open-Sources MiniWorld for Training Video World Models on a Single 8-GPU Server

PekingUniversity · hf · 2026-08-05

Peking University has open-sourced MiniWorld, a lightweight and reproducible framework for training streaming video world models from scratch.

The Problem: Current video world models largely rely on adapting pretrained video generation models via post-training or distillation. These pipelines are complex, compute-heavy, and suffer from mismatches between bidirectional pretraining and causal streaming inference.
Architecture: It 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 stability.
Inference: Combines a rolling KV cache with pipelined asynchronous denoising for efficient streaming generation under bounded computation.
Accessibility: The entire model can be trained end-to-end within several days on a single 8-GPU server. Codebase and pretrained checkpoints have been released to provide a transparent, fully reproducible baseline for the community.

More from Multimodal