When should I use a reference image instead of a better prompt?

Use a reference when the thing you want is easier to show than to describe — a specific face, an exact layout, a palette, an existing frame you only want to change slightly. Use words when you want a new idea rather than a variation of an old one.

There is a simple test for whether to reach for a reference image: try to describe what you want to a person who cannot see it. If your description is getting longer and still not precise — "her nose is like this, the jaw slightly narrower" — you are describing something that words are the wrong tool for. If instead the description is short and clear and the model simply is not delivering it, the problem is the prompt and a reference will not fix it.

Where references genuinely win

GoalReference or promptWhy
Same person across several imagesReferenceIdentity is thousands of tiny proportions; no sentence encodes it
Exact layout or crop you have sketchedReferenceSpatial placement is one of the weakest parts of text conditioning
A specific palette from an existing pictureReference, low strengthColour relationships transfer even when the content does not
Change one element in an image you likeReference / editKeeps everything you did not mention; a rewrite re-rolls the whole frame
A concept you have never seenPromptA reference will anchor you to the thing you already have
Style of a whole genrePromptGenre words are well represented; a single reference over-fits to one picture

Strength is the whole game

Every image-conditioned generation has some notion of how much of the source survives. Low influence gives you a nudge — palette and rough mass, new content. High influence gives you the same picture with a different finish. The frustrating middle is where most disappointing results live: enough of the source to feel derivative, not enough to be consistent. If a result feels like a bad photocopy, move decisively in one direction rather than adjusting by small steps.

What a reference cannot do

  • It will not fix hands or text in the source — those flaws are often preserved or amplified, because the model reads them as intended structure.
  • It will not clean up a low-resolution source into a detailed one on its own; that is an upscaling job, and a different one.
  • It does not override a conflicting prompt. If your words describe a night scene and the reference is daylight, you get an argument, usually won by the reference.

Use references from images you have the right to use. A reference is a much more direct copy of a specific work than a prompt is, and "in the style of" a living artist's specific piece is exactly the case that draws objections.

More on this

Other questions

← all questions