Significant gap in evaluation results of Textdiffuser2

[AnqiGou]

Describe
Model I am using Textdiffuser-2:
I used your layout planner and t2i-full generative model to generate six sub-datasets of MARIOEval with a total of 5k+ images, which were evaluated using the tool for calculating FID and OCR results in Textdiffuser documentation. But I got a FID score of 42.10, which was very different from the results of the paper(33.66). When I tested the OCR results, I used MaskTextSpotterV3 respectively according to the instructions in the repository, and the accuracy of the results was about 32.33, which was very different from the results of the paper(57.58). Same accuracy using pp-ocrv4. What could be the reason? Are there any other notable details?