About

Poetry Illuminated uses GPT-3 and DALL-E-2 to illuminate poems with annotations and images.

You can create up to 10 illuminated poems per month. Click here to get started.

If you have any questions, you can get in touch here.

Limitations

The approach has only been tried on well-known short English-language poems from between the 16th and 20th centuries. Most of these have a regular metre and rhyme scheme. It might fail to produce annotations and images for poems written in other languages and styles or in a different period.

GPT-3 and DALL-E-2 will usually reject any explicit or offensive content but sometimes these filters fail and inappropriate content might be generated. Occasionally failure occurs when innocuous content is considered offensive, for example, a description of a "violent sea" during a storm.

Where it succeeds, the images might not be of a high quality or might not be relevant and annotations might be inaccurate or inadequate. The images may reflect the stereotypes of the time in which the poem was written. They may sometimes reflect biases in the data but not in the poem, for example representing the poet as male even when the poem is written by a woman.

It has difficulty generating images containing several people and tends to distort their faces. It is better with individual people or animals, objects and landscapes.

If you come across any unexpected, surprising or interesting results, please let us know.