HiUni用共享骨干统一视觉理解与生成
AI 总结
论文《HiUni: Harmonizing Understanding and Generation in Shared Visual Space》提出一种统一多模态模型,为视觉理解和生成复用同一套语言对齐的ViT与稠密Transformer骨干网络。模型加入Pixel Diffusion Head进行像素空间生成,并在输出端引入轻量级HiGate通路以保留细粒度空间细节。作者在论文实验中报告,其平均理解得分接近预训练VLM,图像生成和编辑性能具有竞争力。
深度见解
HiUni通过不拆分视觉表征或骨干计算,并以输出端任务专用计算承接差异,为统一多模态建模提供了一种更简洁的架构路径。
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.26107 (cs)
[Submitted on 19 Nov 2026]
Title:HiUni: Harmonizing Understanding and Generation in Shared Visual Space
Authors:Jiahao Hu, Jingdong Du, Yichen Zhang, Yi Song, Zhenyu Zhang, Jiaxin Niu, Ruihang Li, Xiangzhao Hao, Yiqian Zhang, Yong Huang, Zefeng Zhang, Zheyuan Gu, Shuohuan Wang
View a PDF of the paper titled HiUni: Harmonizing Understanding and Generation in Shared Visual Space, by Jiahao Hu and 12 other authors