Description

[TongkunGuan]

How to understand: We employ the HERSHEY font at a size of 10px. On average, one 16x16 patch accommodates approximately 1.5 OPT text tokens. A 224x224 text image contains about 294 text tokens. Consequently, a visual encoder operating on this rendered text image requires only 1/3 of tokens to encode an equivalent amount of text, compared to the text tokenizer in language models.