FLUX.2 dev generations mismatch with diffusers

[bng215]

Hello, I've used your inference script for generating some samples, but got results worse than with diffusers pipeline. It looks like text encoder encodes text in quotes inside prompt in a wrong way or something. It deletes whitespaces between words in english and doesn't work with non-english text inside quotes. Weights are the same.

Here are results of your inference script

And here from diffusers