Skip to lesson
  1. Learn
  2. See
  3. Try
  4. Make

Judge a generated image against its prompt

What this lesson is about

Judge a generated image or clip in four buckets. What was asked, artefacts, style across the set, and fit for the place.

A printer in Meerut shows you a model-made wedding card image. The mangalsutra is drawn as a plain chain. What do you write?

Why this matters

A published study collected rich feedback on eighteen thousand generated images. Raters marked two things separately: the regions that looked wrong, and the prompt words the image missed. The two are kept apart because they are fixed differently. A wrong region can be painted over. A missed word needs a new image, or a better model. That study did not name two more buckets: style across a set, and fit for the place. The listings ask for both.

Lesson 8 of 10 on Multimodal & physical AInextWrite a robot task somebody else can scorePart 3 ends in a work sample you can send

Read the primary or official source: Liang et al., Rich Human Feedback for Text-to-Image Generation (arXiv)