Rich-Human-Feedback 图像数据集

Rich-Human-Feedback 是一个用于文本到图像生成的包含丰富人类反馈的数据集，由 Rapidata 组织于 2025 年发布，旨在为文本到图像生成模型的训练与评估提供丰富反馈，相关论文为「Rich Human Feedback for Text-to-Image Generation」。数据集包含 15k 张图片，它收集了超 15 万人提供的 150 万条注释，包含图像评分、语义一致性、修正建议等反馈。通过多维度的注释方式，获取图像在风格、连贯性、与提示词一致性等方面的评分，标注未对齐单词和图像问题区域，帮助开发者深入理解模型的优势和局限。

Rich-Human-Feedback is a dataset containing rich human feedback for text-to-image generation, released by Rapidata in 2025. It is designed to provide comprehensive human feedback for the training and evaluation of text-to-image generation models, with its associated academic paper titled "Rich Human Feedback for Text-to-Image Generation". The dataset consists of 15,000 images, and includes 1.5 million annotations collected from over 150,000 contributors, covering feedback types such as image quality ratings, semantic consistency assessments, and revision suggestions. Employing multi-dimensional annotation approaches, it collects ratings on aspects including image style, coherence, alignment with the provided prompt, and other relevant metrics, while also marking misaligned terms and problematic regions within the generated images. This resource enables developers to gain in-depth insights into the strengths and limitations of text-to-image generation models.

搜集汇总

背景与挑战

背景概述

Rich-Human-Feedback 图像数据集由 Rapidata 于 2025 年发布，包含 15k 张图片和超过 150 万条人类反馈注释，涵盖图像评分、语义一致性及修正建议等多维度信息。该数据集旨在支持文本到图像生成模型的训练与评估，帮助开发者深入分析模型在风格、连贯性等方面的表现。