DataEvolver introduces a self-evolving multi-agent framework designed to improve text-rich image generation. Unlike traditional static crawl-filter-freeze pipelines, this approach leverages failure signals from rejected samples to iteratively refine data construction. The system aims to enhance the production of visually realistic images with legible, semantically aligned, and layout-consistent text.