Some things are quicker to show than to describe. The exact tilt of a head, a particular shade of washed-out green, the way a room is laid out — you can spend four sentences on any of them and still not land it. Handing the model a picture alongside your text solves in one step what wording often cannot.

But a reference image is not a specification. The model does not know which part of it you cared about. You gave it a photo, and inside that photo are a subject, a pose, a colour palette, a lighting setup, a background and a rendering style, all at once. Unless you say otherwise, it will take an average of the lot, and the result feels vaguely like your reference without matching the thing you actually wanted.

Say what to take and what to ignore

The fix is to name the transfer explicitly in the text. Keep the pose from the reference, change the setting to a winter street. Use the colour palette only, new subject. Same character, different angle. This turns a vague resemblance into a job with edges.

It is equally worth naming what should not come across. If the reference is a phone snapshot and you do not want a phone-snapshot look, say so — otherwise the grain and flat flash light travel along with everything else. Models are literal about what is in front of them, and a scruffy reference produces a scruffy child.

Multiple references need the same discipline, only more of it. Two images with no explanation get blended into a compromise nobody asked for. Two images with roles assigned — this one is the character, this one is the environment — behave much better.

Choose the reference carefully

A good reference is unambiguous. One clear subject, decent resolution, the quality you want visible rather than implied. A busy image with five competing elements gives the model five things to latch onto, and it may not pick yours.

Crop before uploading. If only the jacket matters, send the jacket. Cropping is the most underrated way of saying which part you mean, and it needs no wording at all.

Also, keep expectations sane about likeness. Character consistency across generations is genuinely hard, and even a good reference tends to drift over a series — features soften, proportions shift, small details reinvent themselves. Getting close and accepting a little variation is a healthier goal than chasing an exact double, and re-using your best output as the next reference usually holds a look together better than going back to the original each time.

Text still does the work

A reference narrows the space; it does not fill it. The prompt still has to say what is happening, where, and in what light. People often strip their text down once they attach an image, and then wonder why the result is a slightly worse copy of the reference. Write the prompt as if the image were not there, then add the image to pin down what words handle badly.