Echo-4o-Image
收藏资源简介:
Echo-4o-Image 是一个由 GPT-4o 生成的合成图像数据集,包含约 18 万个样本,涵盖超现实幻想场景、多参考图像生成和复杂指令执行三种任务类型。该数据集旨在解决真实世界图像数据集中存在的不足,如罕见场景和指令对齐问题。通过使用该数据集微调 Bagel 模型,研究人员开发出了新的统一生成模型 Echo-4o,在多个基准测试中表现出色。此外,Echo-4o-Image 还可提升其他基础模型(如 OmniGen2 和 BLIP3-o)的性能,展现了其良好的泛化能力和可迁移性。
Echo-4o-Image is a synthetic image dataset generated by GPT-4o, encompassing approximately 180,000 samples that span three task types: surreal fantasy scenes, multi-reference image generation, and complex instruction execution. Designed to address the deficiencies in real-world image datasets, such as rare scenarios and the alignment of instructions, this dataset has enabled researchers to fine-tune the Bagel model and develop a novel unified generative model, Echo-4o, which has demonstrated exceptional performance in various benchmark tests. Furthermore, Echo-4o-Image has also been shown to enhance the performance of other foundational models (such as OmniGen2 and BLIP3-o), highlighting its strong generalization ability and transferability.
Echo-4o-Image 数据集概述
基本属性
- 任务类别: 文本到图像生成 (text-to-image)
- 语言: 英语 (en)
- 数据规模: 10万至100万条数据之间 (100K<n<1M)
配置信息
- 默认配置名称: default
- 数据文件路径: Surrel-Fantasy-Image/images/0-5000.tar.gz

- 1Echo-4o: Harnessing the Power of GPT-4o Synthetic Images for Improved Image Generation上海人工智能实验室; 中山大学; 香港中文大学多媒体实验室; 北京大学 · 2025年



