Where is the difference lies?

[csgeekhuang]

Congrats for the ICCV2025!

It seems the paper TA-TiTok is very similar to the paper TexTok (https://arxiv.org/pdf/2412.05796) , and I have not seen the citiation in the paper,could you tell me where is the difference? And it is also based on the 1D tokenization(TiTok).

[kdwonn]

Thanks for your interest! This actually came up during the rebuttal as well. The core idea is indeed quite similar to TexTok, but there are a few differences:

1. TA-TiTok leverages text only in the decoder, whereas TexTok uses it in the encoder.
2. TA-TiTok is trained on noisy image–text pairs, while TexTok leverages Gemini-generated captions.

We view these approaches as concurrent works (TA-TiTok was originally submitted to CVPR). I also agree that the manuscript should properly cite TexTok—thanks for pointing that out. I’ll make sure to update it when I get the chance!

[csgeekhuang]

Thanks for your kindly response!