1Frequency-aware 3D RoPE Modulation
Adaptively interpolates under-trained low-frequency components while extrapolating high-frequency components to preserve multi-scale temporal discriminability.
Abstract
Autoregressive video diffusion models have emerged as a scalable paradigm for long video generation. However, they often suffer from severe extrapolation failure, where rapid error accumulation leads to significant temporal degradation when extending beyond training horizons. In this paper, we propose FLEX(Frequency-aware Length EXtension) , a training-free inference-time framework that bridges the gap between short-term training and long-term inference.
On VBench-Long, FLEX outperforms state-of-the-art methods at 6x extrapolation (30s) and remains competitive with long-video fine-tuned baselines at 12x scale (60s). As a plug-and-play augmentation, FLEX extends the generation horizon of existing pipelines (e.g., LongLive) to consistent, dynamic synthesis at minutes level.
Features
2Antiphase Noise Sampling
A structured noise initialization strategy that injects high-frequency dynamic priors to encourage rich temproal variations.
3Inference-only Attention Sink
Adds global structural anchors at inference time for long-range stability, without modifying training or finetuning model weights.
Visual Results
30s Videos (6x Horizon Extension)
Scenes & Objects
CausVid
Self Forcing
LongLive
Rolling Forcing
Self Forcing+FLEX
Sports
CausVid
Self Forcing
LongLive
Rolling Forcing
Self Forcing+FLEX
Portraits
CausVid
Self Forcing
LongLive
Rolling Forcing
Self Forcing+FLEX
Animation
CausVid
Self Forcing
LongLive
Rolling Forcing
Self Forcing+FLEX
60s Videos (12x Horizon Extension)
Self Forcing
LongLive
Rolling Forcing
Self Forcing+FLEX
Ultra-long Video Generation
FLEX is compatible with existing long-video inference pipelines for further generation horizon extension. For instance, it supports more stable multi-minute video generation with LongLive.
Appearance Consistency
Scene Dynamics
LongLive
LongLive+FLEX
LongLive
LongLive+FLEX
Citation
@misc{li2026trainshortinferencelong,
title={Train Short, Inference Long: Training-free Horizon Extension for Autoregressive Video Generation},
author={Jia Li and Xiaomeng Fu and Xurui Peng and Weifeng Chen and Youwei Zheng and Tianyu Zhao and Jiexi Wang and Fangmin Chen and Xing Wang and Hayden Kwok-Hay So},
year={2026},
eprint={2602.14027},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.14027},
}