Text-Render-2M, HQ-Poster-100K, Poster-Preference-100K, Poster-Reflect-120K
收藏资源简介:
PosterCraft是一个用于高质量美学海报生成的统一框架,它通过一系列精心设计的流程优化海报生成。该框架包括四个关键阶段:大规模文本渲染优化、高质量海报微调、美学文本强化学习和视觉语言反馈精炼。为了支持这一流程,我们构建了一套专门的数据集,每个阶段都有自动化的数据构建管道。Text-Render-2M用于文本渲染优化,HQ-Poster-100K包含超过10万张高质量的海报,Poster-Preference-100K生成6000对高质量偏好,Poster-Reflect-120K构建了6.4万对反馈对。这些数据集克服了资源的限制,支持更强大、可迁移的训练,使得训练的模型能够生成高质量的、完整渲染的海报。
PosterCraft is a unified framework for high-quality aesthetic poster generation, which optimizes poster generation via a series of well-designed processes. The framework consists of four key stages: large-scale text rendering optimization, high-quality poster fine-tuning, aesthetic text reinforcement learning, and visual-language feedback refinement. To support this workflow, we have developed a dedicated set of datasets, each equipped with an automated data construction pipeline. Specifically, Text-Render-2M is tailored for text rendering optimization; HQ-Poster-100K contains over 100,000 high-quality posters; Poster-Preference-100K generates 6,000 pairs of high-quality preference samples; and Poster-Reflect-120K constructs 64,000 pairs of feedback pairs. These datasets overcome resource limitations, enabling more powerful and transferable training, and allowing the trained models to generate high-quality, fully rendered posters.



