metadata
license: mit
datasets:
- maxin-cn/SkyTimelapse
- ltzheng/minecraft
language:
- en
base_model:
- facebook/DiT-XL-2-512
- facebook/DiT-XL-2-256
GriDiT: Factorized Grid-Based Diffusion for Efficient Long Image Sequence Generation
Snehal Singh Tomar . Alexandros Graikos . A. Krishna . Dimitris Samaras . Klaus Mueller
Transactions on Machine Learning Research (TMLR) 2026
Stony Brook University
TL;DR: State-of-the-Art image sequence generation models treat image sequences as large tensors of ordered frames. In contrast, our method factorizes image sequence generation into two stages. First, we learn to model the dynamics of the sequence at low resolution, treating the frames as subsampled image grids. Second, we learn to super-resolve individual frames at high resolution. Using the DiT’s self-attention mechanism to model dynamics across frames, and paired with our sampling strategy, our method yields superior synthesis quality for sequences of arbitrary length while significantly reducing sampling time and training data requirements.
Code and Execution Details
Please visit our Github repository.
Citation
Please cite our work as:
@article{
tomar2026gridit,
title={GriDiT: Factorized Grid-Based Diffusion for Efficient Long Image Sequence Generation},
author={Snehal Singh Tomar and Alexandros Graikos and Arjun Krishna and Dimitris Samaras and Klaus Mueller},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=QLD47Ou5lp},
note={}
}