MiniWorld: Democratizing the Training of Video World Models from Scratch
MiniWorld is pitched as a reproducible, from-scratch baseline for streaming video world models that can train on one 8-GPU server.
The paper targets a gap left by heavier approaches that adapt pretrained video generators through post-training or distillation. Its framework uses a block-causal Video Diffusion Transformer with Flow Matching in a pretrained Video VAE latent space. The authors also describe chunk-wise noise scheduling, two-stage continued training, a rolling KV cache, and asynchronous denoising for bounded streaming inference. They say they are releasing code and pretrained checkpoints to make the setup easier to reproduce. HF Daily Papers' note
The paper targets a gap left by heavier approaches that adapt pretrained video generators through post-training or distillation. Its framework uses a block-causal Video Diffusion Transformer with Flow Matching in a pretrained Video VAE latent space. The authors also describe chunk-wise noise scheduling, two-stage continued training, a rolling KV cache, and asynchronous denoising for bounded streaming inference. They say they are releasing code and pretrained checkpoints to make the setup easier to reproduce. HF Daily Papers' note
score 5