A small thread on analysing text rendering capability of #imagen.

Starting with a simple example, Imagen is able to reliably render text without any need for rigorous cherrypicking. e.g. these are non-cherry picked examples for writing "Text-to-Image" on a storefront.

Imagen is also capable of writing text in a variety of
interesting settings. Here are some examples (1 sample picked out of 8 for each prompt).

Some more examples of interesting settings...

As the text gets longer, it naturally becomes more and more difficult to render correctly, but Imagen is pretty good for complex text prompts.
"Text", "Text-to-Image", "Text-to-Image Diffusion", "Text-to-Image Diffusion Models"

The images get worse with very long text renderings, e.g. "Photorealistic Text-to-Image Diffusion Models"
We are still able to find good samples within 8 random generations.

Interestingly, the model is able to write non meaningful combination of letters pretty reliably too. E.g. "magie-xett-ot" (shuffled text-to-image)