NextFlow: Unified Sequential Modeling Activates Multimodal Understanding and Generation
Huichao Zhang, Liao Qu, Yiheng Liu, Hang Chen, Yangyang Song, Yongsheng Dong, Shikun Sun, Xian Li, Xu Wang, Yi Jiang, Hu Ye, Bo Chen, Yiming Gao, Peng Liu, Akide Liu, Zhipeng Yang, Qili Deng, Linjie Xing, Jiyang Liu, Zhao Wang, Yang Zhou, Mingcong Liu
2026-01-06
Summary
This paper introduces NextFlow, a new artificial intelligence model that can understand and generate both text and images at the same time, using a single system.
What's the problem?
Existing AI models often treat text and images separately, requiring different systems for each. Generating high-quality images with these models can also be very slow, and creating images at different scales (like zooming in or out) can be unstable. The way images are traditionally created by AI, scanning them line by line, isn't the most efficient way to represent how we actually *see* images, which is more hierarchical.
What's the solution?
The researchers created NextFlow, which uses a single 'brain' to process both text and images. For text, it predicts the next word in a sequence, which is standard. But for images, it predicts the next level of detail – essentially building the image from broad shapes to finer details, instead of scanning it. This 'next-scale' prediction makes image generation much faster, creating 1024x1024 images in just 5 seconds. They also developed techniques to stabilize the image generation process and a way to fine-tune the model using reinforcement learning.
Why it matters?
NextFlow represents a significant step forward in AI because it can handle multiple types of data with one model, leading to more versatile applications like image editing, creating content that mixes text and images, and even generating videos. It's also much faster and produces images of comparable quality to specialized AI models, making it a more practical solution for many tasks.
Abstract
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.