IRGL-300K
收藏资源简介:
IRGL-300K是一个包含300,000条数据的数据集,由华东师范大学、香港中文大学、小红书、加州大学洛杉矶分校、浙江大学、牛津大学的研究人员共同创建。该数据集旨在用于训练一个名为Interleaving Reasoning Generation (IRG)的文本到图像生成模型。数据集包括六个分解的学习模式,共同覆盖了基于文本的思考和完整的思考-图像轨迹。数据集的创建过程包括两个阶段的训练,首先建立强大的思考和反思,然后利用完整的思考-图像轨迹数据有效地调整IRG流程。IRG在多个基准测试中取得了最先进的性能,并在视觉质量和细粒度保真度方面取得了显著的改进。
IRGL-300K is a dataset consisting of 300,000 instances, jointly created by researchers from East China Normal University, The Chinese University of Hong Kong, Xiaohongshu, University of California, Los Angeles, Zhejiang University, and the University of Oxford. This dataset is intended for training a text-to-image generation model named Interleaving Reasoning Generation (IRG). The dataset includes six decomposed learning paradigms that collectively cover text-based thinking and complete thinking-to-image trajectories. The dataset creation process involves two-stage training: first establishing robust thinking and reflection capabilities, then leveraging complete thinking-to-image trajectory data to effectively fine-tune the IRG pipeline. IRG has achieved state-of-the-art performance across multiple benchmarks, with notable improvements in visual quality and fine-grained fidelity.
数据集概述
基本信息
- 数据集名称:Interleaving Reasoning Generation
- 官方仓库:https://github.com/Osilly/Interleaving-Reasoning-Generation
- 关联论文:Interleaving Reasoning for Better Text-to-Image Generation(https://arxiv.org/abs/2509.06945)
性能表现
| 模型 | GenEval | WISE | TIIF-short/long (Qwen) | GenAI-Bench | OneIG-EN |
|---|---|---|---|---|---|
| Janus-Pro-7B | 0.80 | 0.35 | 65.38/61.10 | 0.75 | 0.267 |
| FLUX.1-dev | 72.9 | 0.50 | 66.24/66.72 | 0.76 | 0.434 |
| Show-o2 | 0.76 | 0.61 | 62.80/63.87 | 0.75 | 0.308 |
| BAGEL | 0.78 | 0.52 | 70.97/71.79 | 0.79 | 0.361 |
| BAGEL w/ self-CoT | 0.79 | 0.70 | 68.06/68.78 | 0.81 | 0.324 |
| IRG | 0.85 | 0.77 | 76.00/73.77 | 0.84 | 0.415 |
方法概述
- 核心思想:交错推理生成(Interleaving Reasoning Generation,IRG)
- 流程:给定提示词,模型首先生成基于文本的推理过程,然后基于该推理生成图像。接着,在初始图像的基础上,模型反思如何改进质量,并通过反思过程生成优化后的图像。
- 优势:通过多轮推理显著提升图像生成质量,包括增强渲染纹理、阴影真实感等视觉属性,以及改进精细细节(如手指结构)。
时间线
- 2025年9月8日:IRG论文(Interleaving Reasoning for Better Text-to-Image Generation)发布于arXiv。
状态说明
- 数据集、代码和权重尚未发布,请持续关注。




