_version_ 1866912802637086720
author Zhang, Huichao
Qu, Liao
Liu, Yiheng
Chen, Hang
Song, Yangyang
Dong, Yongsheng
Sun, Shikun
Li, Xian
Wang, Xu
Jiang, Yi
Ye, Hu
Chen, Bo
Gao, Yiming
Liu, Peng
Liu, Akide
Yang, Zhipeng
Deng, Qili
Xing, Linjie
Liu, Jiyang
Wang, Zhao
Zhou, Yang
Liu, Mingcong
Zhang, Yi
He, Qian
Hu, Xiwei
Qi, Zhongqi
Shao, Jie
Fu, Zhiye
Wang, Shuai
Chen, Fangmin
Chai, Xuezhi
Wu, Zhihua
Wang, Yitong
Yuan, Zehuan
Du, Daniel K.
Wu, Xinglong
author_facet Zhang, Huichao
Qu, Liao
Liu, Yiheng
Chen, Hang
Song, Yangyang
Dong, Yongsheng
Sun, Shikun
Li, Xian
Wang, Xu
Jiang, Yi
Ye, Hu
Chen, Bo
Gao, Yiming
Liu, Peng
Liu, Akide
Yang, Zhipeng
Deng, Qili
Xing, Linjie
Liu, Jiyang
Wang, Zhao
Zhou, Yang
Liu, Mingcong
Zhang, Yi
He, Qian
Hu, Xiwei
Qi, Zhongqi
Shao, Jie
Fu, Zhiye
Wang, Shuai
Chen, Fangmin
Chai, Xuezhi
Wu, Zhihua
Wang, Yitong
Yuan, Zehuan
Du, Daniel K.
Wu, Xinglong
contents We present NextFlow, a unified decoder-only autoregressive transformer trained on 6 trillion interleaved text-image discrete tokens. By leveraging a unified vision representation within a unified autoregressive architecture, NextFlow natively activates multimodal understanding and generation capabilities, unlocking abilities of image editing, interleaved content and video generation. Motivated by the distinct nature of modalities - where text is strictly sequential and images are inherently hierarchical - we retain next-token prediction for text but adopt next-scale prediction for visual generation. This departs from traditional raster-scan methods, enabling the generation of 1024x1024 images in just 5 seconds - orders of magnitude faster than comparable AR models. We address the instabilities of multi-scale generation through a robust training recipe. Furthermore, we introduce a prefix-tuning strategy for reinforcement learning. Experiments demonstrate that NextFlow achieves state-of-the-art performance among unified models and rivals specialized diffusion baselines in visual quality.
format Preprint
id arxiv_https___arxiv_org_abs_2601_02204
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle NextFlow: Unified Sequential Modeling Activates Multimodal Understanding and Generation
Zhang, Huichao
Qu, Liao
Liu, Yiheng
Chen, Hang
Song, Yangyang
Dong, Yongsheng
Sun, Shikun
Li, Xian
Wang, Xu
Jiang, Yi
Ye, Hu
Chen, Bo
Gao, Yiming
Liu, Peng
Liu, Akide
Yang, Zhipeng
Deng, Qili
Xing, Linjie
Liu, Jiyang
Wang, Zhao
Zhou, Yang
Liu, Mingcong
Zhang, Yi
He, Qian
Hu, Xiwei
Qi, Zhongqi
Shao, Jie
Fu, Zhiye
Wang, Shuai
Chen, Fangmin
Chai, Xuezhi
Wu, Zhihua
Wang, Yitong
Yuan, Zehuan
Du, Daniel K.
Wu, Xinglong
Computer Vision and Pattern Recognition
Artificial Intelligence
We present NextFlow, a unified decoder-only autoregressive transformer trained on 6 trillion interleaved text-image discrete tokens. By leveraging a unified vision representation within a unified autoregressive architecture, NextFlow natively activates multimodal understanding and generation capabilities, unlocking abilities of image editing, interleaved content and video generation. Motivated by the distinct nature of modalities - where text is strictly sequential and images are inherently hierarchical - we retain next-token prediction for text but adopt next-scale prediction for visual generation. This departs from traditional raster-scan methods, enabling the generation of 1024x1024 images in just 5 seconds - orders of magnitude faster than comparable AR models. We address the instabilities of multi-scale generation through a robust training recipe. Furthermore, we introduce a prefix-tuning strategy for reinforcement learning. Experiments demonstrate that NextFlow achieves state-of-the-art performance among unified models and rivals specialized diffusion baselines in visual quality.
title NextFlow: Unified Sequential Modeling Activates Multimodal Understanding and Generation
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2601.02204