Question about the training dataset - tokenzier

[Lucky-Light-Sun]

Hi, I recently read your paper “Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction” published in CVPR 2024. I am very interested in your method and have great respect for your contribution to the academic community.

We are currently doing some experiments and would like to ask you a question about the tokenizer part. Is the pre-training dataset you are using OpenImage V6 or another version? Current we are using OpenImage v6 to train tokenizer. So do you think it's fair to compare with VAR tokenizer?

Thank you again for your time and effort, and I look forward to your reply.