Efstathios Karypidis1,3, Spyros Gidaris2, Nikos Komodakis1,4,5
1Archimedes/Athena RC 2valeo.ai
3National Technical University of Athens 4University of Crete 5IACM-Forth
Code and pretrained models will be released soon. Stay tuned!
This framework is built on top of Cosmos-Predict2 and DINO-Foresight. Cosmos-Predict is a collection of general-purpose world foundation models. The repository contains many features that are not used in Re2Pix. We did an initial cleanup of the repository, but it still contains a lot of unused code. To be refined further in the future.
For the installation instructions, please refer to the Installation page.
Re2Pix is a hierarchical video prediction framework that takes as input 13 frames of a video and predicts the next 12 frames in a hierarchical manner. Resolution used is 448x. For instructions on how to use Re2Pix, please refer to the Usage page.
- Provide inference script
- Add training scripts
- Improve the documentation and code cleanup
- Add more pretrained models
- 2026-06-17: Re2Pix is accepted at ECCV 2026! 🎉
- 2026-04-13: Arxiv Preprint and GitHub repository are released!
If you found Re2Pix useful in your research, please consider starring ⭐ us on GitHub and citing 📚 us in your research!
@article{karypidis2026re2pix,
title={Representations Before Pixels: Semantics-Guided Hierarchical Video Prediction},
author={Karypidis, Efstathios and Gidaris, Spyros and Komodakis, Nikos},
journal={arXiv preprint arXiv:2604.11707},
year={2026}
}