August 5, 2026 MiniWorld: Democratizing the Training of Video World Models from Scratch MiniWorld 提出轻量级可复现的视频世界模型训练框架，基于块因果 Video DiT + Flow Matching + 块级非递减噪声调度 + 滚动 KV Cache 流水线异步去噪，可在单台 8-GPU 服务器上数天内完成训练 video-world-model diffusion-transformer flow-matching block-causal-attention streaming-inference embodied-ai diffusion-forcing kv-cache