# FreeNoise: Tuning-Free Longer Video Diffusion Via Noise Rescheduling

**Authors:** Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, Ziwei Liu  
**arXiv:** [2310.15169](https://arxiv.org/abs/2310.15169)  
**Date:** 2023-10-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.15169v1/

## Key points

- Existing video generation models have limitations in terms of the number of frames they can be trained on and their support for single-text conditions.
- The authors propose a method called FreeNoise to enhance the generative capabilities of pretrained video diffusion models.
- FreeNoise is a tuning-free and time-efficient paradigm that reschedules noises for long-range correlation and applies temporal attention using a window-based function.
- A novel motion injection method is introduced to support the generation of videos conditioned on multiple text prompts.
- FreeNoise outperforms previous methods in extending generative capabilities with negligible time cost (approximately 17%).
- Generated video samples are available on the authors' website for further evaluation.

## Summary

This study addresses the limitations of existing video generation models, which are typically trained on a limited number of frames and only support single-text conditions. To overcome these challenges, the authors propose a method called FreeNoise to extend the generative capabilities of pretrained video diffusion models. The first contribution of this study is an analysis of the impact of initial noise in video diffusion models. Based on this observation, the authors introduce FreeNoise as a tuning-free and time-efficient paradigm to enhance the generative capabilities while preserving content consistency. Instead of initializing noises for all frames, FreeNoise reschedules a sequence of noises for long-range correlation and applies temporal attention over them using a window-based function. The second contribution is a novel motion injection method designed to support the generation of videos conditioned on multiple text prompts. This addresses real-life scenarios where video content changes over time and requires multi-text conditions. Extensive experiments validate the superiority of FreeNoise in extending the generative capabilities of video diffusion models. Notably, compared to previous methods that incurred significant extra time costs (255%), FreeNoise incurs only negligible time cost (approximately 17%). The authors provide generated video samples on their website for further evaluation. Overall, this study presents an innovative approach to text-driven video generation by extending model capabilities to generate longer videos conditioned on multiple texts. The proposed FreeNoise paradigm demonstrates improved performance while maintaining efficiency, making it an invaluable contribution to the field.