arXiv 2504.029492025-04-03cs.CVcs.AI

VARGPT-v1.1：通过迭代指令微调与强化学习改进视觉自回归大型统一模型

VARGPT-v1.1: Improve Visual Autoregressive Large Unified Model via Iterative Instruction Tuning and Reinforcement Learning

School of Electronic and Computer Engineering, Peking University（北京大学电子与计算机工程学院）；The Chinese University of Hong Kong（香港中文大学）

Xianwei Zhuang, Yuxin Xie, Yufan Deng, Dongchao Yang, Liming Liang, Jinghan Ru, Yuguo Yin, Yuexian Zou

备注：: Code is available at: https://github.com/VARGPT-family/VARGPT-v1.1. arXiv admin note: text overlap with arXiv:2501.12327

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

VARGPT-v1.1通过结合迭代视觉指令微调与DPO强化学习，采用Qwen2骨干并扩展数据集，在保持架构一致性的同时实现了图像编辑能力，显著提升了多模态理解与文本到图像生成的性能。

摘要：

在这项工作中，我们提出了VARGPT-v1.1，这是一种先进的统一视觉自回归模型，构建在我们之前的框架VARGPT之上。该模型保留了用于视觉理解的下一词元预测和用于图像合成的下一尺度生成的双重范式。具体而言，VARGPT-v1.1集成了：(1)一种结合迭代视觉指令微调与通过直接偏好优化（DPO）进行强化学习的新型训练策略；(2)包含8.3M个视觉生成指令对的扩展训练语料库；(3)使用Qwen2升级的语言模型骨干；(4)增强的图像生成分辨率；(5)无需架构修改即可实现的涌现图像编辑能力。这些进步使VARGPT-v1.1在多模态理解和文本到图像指令遵循任务中实现了最先进的性能，在理解和生成指标上均展现出显著提升。值得注意的是，通过视觉指令微调，该模型获得了图像编辑功能，同时保持了与前一版本的架构一致性，揭示了统一视觉理解、生成和编辑的潜力。我们的研究结果表明，设计良好的统一视觉自回归模型可以有效采用大型语言模型（LLMs）的灵活训练策略，展现出良好的可扩展性。代码库和模型权重已在https://github.com/VARGPT-family/VARGPT-v1.1公开提供。

英文摘要：

In this work, we present VARGPT-v1.1, an advanced unified visual autoregressive model that builds upon our previous framework VARGPT. The model preserves the dual paradigm of next-token prediction for visual understanding and next-scale generation for image synthesis. Specifically, VARGPT-v1.1 integrates: (1) a novel training strategy combining iterative visual instruction tuning with reinforcement learning through Direct Preference Optimization (DPO), (2) an expanded training corpus containing 8.3M visual-generative instruction pairs, (3) an upgraded language model backbone using Qwen2, (4) enhanced image generation resolution, and (5) emergent image editing capabilities without architectural modifications. These advancements enable VARGPT-v1.1 to achieve state-of-the-art performance in multimodal understanding and text-to-image instruction-following tasks, demonstrating significant improvements in both comprehension and generation metrics. Notably, through visual instruction tuning, the model acquires image editing functionality while maintaining architectural consistency with its predecessor, revealing the potential for unified visual understanding, generation, and editing. Our findings suggest that well-designed unified visual autoregressive models can effectively adopt flexible training strategies from large language models (LLMs), exhibiting promising scalability. The codebase and model weights are publicly available at https://github.com/VARGPT-family/VARGPT-v1.1.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。