Karine-Huang / T2I-CompBench

[Neurips 2023] T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation
https://arxiv.org/pdf/2307.06350.pdf
MIT License
213 stars 7 forks source link

Making sure to run the benchmark correctly #17

Closed royg27 closed 5 months ago

royg27 commented 5 months ago

Hello,

First of all, thanks for the interesting work and the useful benchmark!

I have a new text-to-image generative model, and I want to make sure that I evaluate and compare it properly (as I aim to report it in a paper): First, I generate the 300 samples from texture_val.txt along with their index. Next, I ran the BLIP_vqa script.

I saw that running with and without the index number (caption_index.png) significantly changes the results. Should it be this way? If so, the index should be set according to the order in the txt files?

Thanks a lot!