How to improve typeface rendering in image generators (specifically autoregressive and AR+diffusion based models)? Is this done with datasets of typeface specimens (both of open source & commercial typefaces) with associated captions (to feed into the generator's language models) to improve the text rendering, in order to make the rendered text resemble the actual typeface (not necessarily exactly through), without needing dedicated font files to do so? (The generated text may also suffer from inconsistent spacing and etc. but this isn't a big problem).

Full article content could not be extracted automatically. Read the original below.