GenRef
收藏资源简介:
GenRef是一个大规模的图像反思数据集,由香港中文大学MMLab创建,包含100万个反思三元组,每个三元组由一个反思、一个有缺陷的图像和一个增强的图像组成。该数据集通过四个不同的数据来源,利用可验证的目标、集成奖励模型和多样化的滚动策略,构建了一个自动化的数据构建管道。GenRef旨在为图像生成模型提供自反思能力,通过迭代评估和纠正之前生成的输出,从而提高图像质量。
GenRef is a large-scale image reflection dataset created by the MMLab of The Chinese University of Hong Kong. It contains 1 million reflection triplets, each consisting of a reflection annotation, a defective image, and an enhanced image. This dataset constructs an automated data construction pipeline through four distinct data sources, leveraging verifiable objectives, integrated reward models, and diverse rolling strategies. GenRef aims to equip image generation models with self-reflection capabilities, improving image quality via iterative evaluation and correction of previously generated outputs.

- 1From Reflection to Perfection: Scaling Inference-Time Optimization for Text-to-Image Diffusion Models via Reflection Tuning香港中文大学 · 2025年



