When Pretty Isn't Useful: Investigating Why Modern Text-to-Image Models Fail as Reliable Training Data Generators
当漂亮并不有用:调查现代文本到图像模型为何无法作为可靠的训练数据生成器
机构 * RPTU University Kaiserslautern-Landau(凯撒斯劳滕-兰道大学) ; German Research Center for Artificial Intelligence (DFKI)(德国人工智能研究中心)
专题命中 文生图 :text-to-image(title,abstract);diffusion(abstract);分类 cs.CV
AI总结 研究发现现代文本到图像模型在生成合成数据时性能下降,揭示其生成的数据分布狭窄且缺乏多样性,挑战了生成现实感进步即意味着数据现实感进步的假设。
Comments Accepted to CVPR26, Project Page: https://bill2462.github.io/When-Pretty-Isn-t-Useful-page/, Code: https://github.com/Bill2462/When-Pretty-Isn-t-Useful-codebase