Color is the first thing anyone notices in a generated image and the last thing most people bother to describe. Type “a cafe interior” and the model picks a palette for you, usually a warm brown one, because that is what the average cafe photo looks like. If you want something else, you have to say so — but the way you say it matters more than how much you say.

The common mistake is listing colors like a paint order: red chairs, blue walls, yellow lamps, green plants. Image models do not assign colors to objects the way a form assigns values to fields. Color words leak. Ask for a blue wall and a red chair and you may well get a purple chair standing against a bluish wall, because the model is blending descriptions rather than filling slots. The more separate color instructions you stack, the more they bleed into each other.

Describe the palette, not every object

A far more reliable move is to name the overall palette in one phrase and let the model distribute it. “Muted earth tones,” “cool blue-grey palette,” “high-contrast black and orange,” “washed-out pastels” — each of these gives the model a coherent target instead of five competing ones. Then pick one object you actually care about and give it a color. One exception inside a stated palette usually survives; five exceptions do not.

Borrowed vocabulary works well here. Terms from photography and design carry a lot of color information in very few words: golden hour, overcast, neon-lit, candlelit, fluorescent office light, moonlight. These describe light rather than paint, and because light governs every color in the scene, they shift the whole image at once and stay internally consistent. “A cafe interior at golden hour” beats a list of six color adjectives almost every time.

Temperature, saturation, contrast

If you want finer control without listing objects, think in three dials. Temperature is warm versus cool — say it plainly. Saturation runs from muted, desaturated and faded through to vivid and punchy. Contrast goes from soft and flat at one end to harsh and high-contrast at the other. Naming even two of these three usually gets you closer than a paragraph of color nouns, and they combine cleanly because they describe treatment rather than content.

Reference styles are the shortcut when you know one. Technicolor, cyanotype, risograph print, faded Polaroid, old film stock — each implies a whole palette plus a grain and a contrast curve. Use them when the look you want already has a name, but be aware they bring extra baggage: a risograph reference will also flatten your shapes and coarsen the detail.

When color will not stick

Some colors simply resist. Very specific shades — a particular teal, a certain brick red — rarely land exactly, because the model has no color picker, only associations with words. If a shade genuinely matters, describe it through an object that already has it: “the green of an old chalkboard” works better than “dark muted green.” And accept that pure white and pure black backgrounds tend to drift grey unless you ask for them directly.

The practical habit is to write the prompt without color first, get the composition right, then add one palette phrase and one lighting phrase. Two lines of color language applied to a working image will do more than a color list applied to a broken one.