AtlasVid

Efficient Ultra-High-Resolution Long Video Generation
via Decoupled Global-Local Modeling

NeurIPS 2026

Ziyang Mai*Yuyao Zhang*Yu-Wing Tai

Dartmouth College

* Equal contribution

Speed, method and results in four minutesWatch on YouTube

Overview

AtlasVid separates global scene structure from local visual detail. A low-resolution, low-frame-rate semantic proxy guides a detail generator through hierarchical, locality-preserving attention. Trained at 720p with lightweight LoRA adaptation, the model scales to 4K and beyond without training on ultra-high-resolution long videos.

Temporally scaled RoPE extends the generation horizon, while aligned global guidance maintains coherence throughout a continuous shot.

Runtime

4K · 81 frames

Left: runtime to generate one 4K, 81-frame video on a log scale. UltraWan takes about 29 hours, UltraGen 11,536 seconds (3.2 hours), T3-Video 2,544 seconds (42.4 minutes), and AtlasVid 1,715 seconds (28.6 minutes), 60.9 times faster than UltraWan. Right: attention FLOPs per layer from 720p to 8K at 81 frames for FlashAttention, UltraGen, T3-Video, linear attention and AtlasVid; at 8K AtlasVid needs 1208.2 times fewer FLOPs than FlashAttention.
Runtime and attention FLOPs. Figure 1 in the paper ↗

More results

4K source · 161 frames

Snow-covered forestView full size
Ocean wavesView full size
Crab and sea anemoneView full size
Sunlit interiorView full size

Videos are compressed to 720p for web playback. Source resolutions are indicated above.

Method

A global semantic proxy establishes the scene and motion. The detail branch jointly denoises high-resolution tokens using local attention and aligned global guidance.

AtlasVid pipeline: a global semantic generator guides spatiotemporal detail generation, with locality-preserving attention, scaled RoPE, and resolution-agnostic training.
AtlasVid framework. Figure 2 in the paper ↗

Ablation studies

Global guidance and locality-preserving attention

Global guidance maintains scene consistency, while locality-preserving attention enables resolution extrapolation. All variants are trained at 720p and evaluated at 4K. For the locality comparison, attention between neighboring blocks is disabled to expose the block structure.

Iceberg generation comparison: the full model, without global guidance, with locality-preserving attention, and without locality-preserving attention.
Attention design ablation. Figure 4 in the paper ↗

Optional 4K fine-tuning

The base model already generates plausible 4K detail. An optional refinement stage using short 4K clips further improves local realism, as illustrated by the brick textures.

A brick pier with optional 4K fine-tuning on top and without 4K fine-tuning below, with enlarged texture details.
4K fine-tuning ablation. Figure 5 in the paper ↗

BibTeX

@inproceedings{mai2026atlasvid,
  title={AtlasVid: Efficient Ultra-High-Resolution Long Video
         Generation via Decoupled Global-Local Modeling},
  author={Mai, Ziyang and Zhang, Yuyao and Tai, Yu-Wing},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}