# MilliVid Uses Hierarchical Latents to Maintain Consistency Across 256+ Video Frames

_Research · published 2026-06-11_

Researchers have introduced MilliVid, a video generation approach that tackles long-range consistency by structuring the latent space hierarchically across multiple resolutions. Rather than forcing a transformer to handle ever-longer token sequences, MilliVid trains a hierarchical autoencoder that compresses each frame into coarse-to-fine token levels — from full-resolution grids down to just a handful of tokens per frame — and then runs a diffusion model over these levels using a coarse-to-fine rollout strategy. The coarsest levels encode scene layout and semantics, while finer levels carry texture and high-frequency detail, letting the model prioritize compute where it matters most for perceptual consistency.

Validated on a custom dataset of long Minecraft videos, MilliVid substantially outperforms FramePack and standard autoregressive rollout on both consistency and quality metrics across up to 256 frames of context. Crucially, the baselines degrade below random-frame consistency at longer rollout lengths due to exposure bias, while MilliVid remains stable — a result that has clear implications for any generative video application requiring persistent scene geometry and object permanence over extended sequences.

## Sources
- [davidcharatan.com](https://davidcharatan.com/millivid/)
---
Canonical: https://genbuzz.news/posts/millivid-uses-hierarchical-latents-to-maintain-consistency-across-256-video-frames