Reproduce “Fine-T2I: An Open, Large-Scale, and Diverse Dataset for High-Quality T2I Fine-Tuning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xu Ma, Yitian Zhang, Qihua Dong, Yun Fu

Northeastern University researchers introduce Fine-T2I, an open, large-scale dataset featuring over 6 million high-quality text-image pairs meticulously curated for Text-to-Image (T2I) fine-tuning. Models fine-tuned on Fine-T2I demonstrate enhanced visual quality and text-image alignment, achieving up to an 80.7% win rate for visual quality in human evaluations and outperforming existing open-source datasets.

Key insights

  • A meticulous, multi-stage data curation pipeline, including VLM-powered verification for visual quality and text-image alignment, is crucial for building effective T2I fine-tuning datasets.
  • High-quality synthetic data, generated with controlled prompt design and rigorous filtering, can effectively bridge data gaps and significantly enhance the diversity and scale of T2I training resources.
  • Providing both original and enhanced prompts within the dataset helps train T2I models to better interpret and generate images from varied user inputs, from concise to detailed descriptions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.