Drop rates

Hi, guys! In your paper you write that you drop text, vit and vae tokens with probs 0.1, 0.5 and 0.1. Is this true, or this is a typo, and you dropped vae tokens with 0.5 rate?
I am now trying to make a pretrain with these settings, and it looks like model is overconditioned on vae tokens, and without them (on t2i) quality degrades during train.

Compare the paper's stated text, ViT, and VAE token-drop probabilities (0.1, 0.5, and 0.1) with the pretraining settings used by Bagel. Confirm whether the VAE rate is 0.1 or 0.5, then document the authoritative value and its effect on the t2i setup.

Written by the indexing model from the issue text.