OpenUni Training Dataset
收藏资源简介:
OpenUni 数据集是一个开放源代码的多模态理解和生成框架,由南洋理工大学 S-Lab, 商汤科技 SenseTime Research, 商汤科技和 Tetras.AI 的研究人员共同创建。该数据集包含约 2300 万个图像-文本对,用于训练和微调 OpenUni 模型,使其能够在图像生成和理解任务上取得优异的性能。数据集的创建过程采用了两阶段训练策略,首先在预训练阶段将大型语言模型和多模态模型进行对齐,然后在微调阶段使用高质量的数据进行优化。OpenUni 数据集的应用领域包括图像生成、多模态理解和世界知识理解,旨在解决多模态人工智能中的理解和生成任务统一问题。
The OpenUni Dataset is an open-source multimodal understanding and generation framework, co-created by researchers from S-Lab at Nanyang Technological University, SenseTime Research, SenseTime Group, and Tetras.AI. This dataset contains approximately 23 million image-text pairs, which are utilized to train and fine-tune the OpenUni model to achieve outstanding performance on image generation and understanding tasks. The dataset's creation process adopts a two-stage training strategy: first, aligning large language models with multimodal models during the pre-training phase, and then optimizing with high-quality data during the fine-tuning phase. The application domains of the OpenUni Dataset include image generation, multimodal understanding, and world knowledge understanding, aiming to resolve the unified challenge of understanding and generation tasks in multimodal artificial intelligence.
OpenUni数据集概述
基本信息
- 名称: OpenUni
- 类型: 多模态理解与生成基准模型
- 开发者: Size Wu, Zhonghua Wu, Zerui Gong, Qingyi Tao, Sheng Jin, Qinyue Li, Wei Li, Chen Change Loy
- 论文地址: https://arxiv.org/abs/2505.23661
- 许可证: NTU S-Lab License 1.0
模型变体
| 模型名称 | 图像尺寸 | MLMM模型 | 扩散模型 | 预训练权重 | 微调权重 |
|---|---|---|---|---|---|
| OpenUni-B-512 | 512×512 | InternVL3-1B | SANA-0.6B-512px | 预训练链接 | 微调链接 |
| OpenUni-L-512 | 512×512 | InternVL3-2B | SANA-1.6B-512px | 预训练链接 | 微调链接 |
| OpenUni-L-1024 | 1024×1024 | InternVL3-2B | SANA1.5-1.6B-1024px | 预训练链接 | 微调链接 |
主要功能
- 生成高质量且符合指令的图像
- 在标准基准测试(GenEval, DPG-Bench, WISE)上表现优异
技术特点
- 激活参数仅1.1B和3.1B
- 基于InternVL3作为多模态LLM基础
- 使用SANA作为扩散模块
使用说明
-
权重下载: 通过HuggingFace Hub下载 bash huggingface-cli download wusize/openuni --local-dir checkpoints --repo-type model
-
推理指南: 参考docs/INFERENCE.md
-
评估方法: 参考docs/EVALUATION.md
致谢
- SANA
- InternVL3
- MetaQuery
- BLIP3-o团队

- 1OpenUni: A Simple Baseline for Unified Multimodal Understanding and Generation南洋理工大学 S-Lab, 商汤科技 SenseTime Research, 商汤科技和 Tetras.AI · 2025年



