search

YangLing0818 / IterComp

IterComp: Iterative Composition-Aware Feedback Learning from Model Gallery for Text-to-Image Generation
https://arxiv.org/abs/2410.07171
MIT License
129 stars 8 forks source link

Recommending our work on evaluation of T2I models! #2

Closed linzhiqiu closed 2 weeks ago

linzhiqiu commented 3 weeks ago

Great work! I also wanted to advertise our recent work on evaluating compositional generation, which might be of interest to you:

VQAScore: VQAScore is a simple but effective automated scoring metric for image-text alignment. VQAScore strongly agrees with human judgments on compositional text prompts and can be run using one-line Python code in our repo here! It also significantly outperforms previous automated metrics such as ImageReward, HPSv2, and PickScore. Google's Imagen3 reported VQAScore as the strongest replacement for CLIPScore. GenAI-Bench: We introduce a comprehensive benchmark for compositional image generation with 1,600 prompts written by professional designers. We also show that VQAScore can be used as a strong reward metric by re-ranking the generated candidate images. GenAI-Bench was presented as the Best Short Paper at SynData@CVPR24 workshop and was highlighted in Imagen3's report.

Cominclip commented 3 weeks ago

Thank you for your recommendation! Both of the works you suggested are highly valuable and solid. We will use them to conduct a more comprehensive evaluation of IterComp. We will also keep following these two works and look forward to your future high-quality research. Thank you for your interest in IterComp!

linzhiqiu commented 3 weeks ago

Thanks! Looking forward to your update and future work too!