Color Prompts for AI Images: Do Hex Codes Work?

Quick answer: hex codes work in AI image prompts on some models only. FLUX.2 documents them; Ideogram says it does not read hex in prompt text outside its JSON mode on 4.0; Google's and OpenAI's guides do not use them. Named colors work everywhere, so lead with a hue plus a qualifier, and tie any hex code to one named object.

Naming a color and getting that color are different problems, and the gap is vendor-specific. This page reports what each vendor documents, not generated test images: I read the documentation rather than fabricating comparisons.

Last verified October 2026: on 27 October 2026 I re-checked FLUX.2's hex guidance at docs.bfl.ai and Ideogram's statement that it does not understand numerical color codes in prompt text. The other vendor claims below were read in September 2026 and not re-checked.

Jump to: hex or names? · Ideogram palette setting · color words and modifiers · palette in one line · reference images · grading · copy-paste prompts. For brand-wide color rules across a whole image set, see brand-consistent AI imagery; this page covers specifying a color in one prompt.

Which Models Accept Hex Codes? Quick Answer

Read the middle row carefully before assuming any vendor besides Ideogram has a color-specific control at all. Nano Banana Pro's request shape, response_format with aspect_ratio and image_size, has no color field of any kind; color only ever enters through the words in the prompt itself. GPT Image 2's request shape is the same story: size, quality and a handful of formatting options, nothing that names a color channel. Ideogram is the outlier twice over, once for explicitly ruling hex out of prompt text, and again for being the only one of the four with a discrete, non-text color control at all.

Why Doesn't Naming a Color Guarantee You Get It?

Because a color word is relational, not a fixed target the way a hex code is on a screen. "Red" covers a wide band from brick to cherry to fire-engine, and a model trained on millions of captioned photos learned "red" as whatever range of pixels people tended to call red, not as one coordinate. A hex code looks like it should sidestep that ambiguity entirely, since it names one exact value with no room for interpretation. Whether that promise actually holds turns out to depend on which model you're asking, checked below against each vendor's own current documentation rather than assumed to work the same way everywhere.

Do Hex Codes Actually Work in AI Image Prompts?

Genuinely, model by model, not as one universal answer.

FLUX.2 documents it directly, with worked examples. Black Forest Labs' own guide states plainly: "FLUX.2 supports precise color matching using hex codes." Its instructions say to signal hex colors with a keyword like color or hex followed by the code, and to associate each hex value clearly with a specific object or surface. A published worked example reads: "a vintage illustration of an apple in color #0047AB with a heart-shaped cutout in the middle, on a white background". The same guide adds an honest limit rather than an unqualified promise: "Hex codes work best when clearly associated with specific objects." A vague reference naming a color with no object attached, its own example is asking for a color "somewhere" in the image, is called out as producing inconsistent results.

Ideogram's own docs say the opposite, for prompt text. Its prompting guide states: "Ideogram doesn’t understand numerical color codes (like RGB or hex)" in a typed prompt, unless the request goes through its structured JSON prompting mode on Ideogram 4.0 (docs.ideogram.ai, accessed September 2026). Its recommended alternative is naming colors the way a person would describe them from memory, more on that below.

Neither OpenAI's nor Google's current image-prompting guides mention hex codes at all. Checked directly against platform.openai.com/docs/guides/image-generation, OpenAI's cookbook prompting guide, and Google's Gemini image-generation guide: zero genuine occurrences of "hex" as color guidance on any of the three (a handful of unrelated "hex" hits exist in each page's own JavaScript, not in the documentation text). Both vendors' own worked examples model color entirely through natural language instead, covered next.

What's Different About Ideogram's Color Palette Setting?

Ideogram has a second, entirely separate color mechanism that isn't prompt text at all: a Color Palette setting, a discrete field in the generation settings, not something parsed out of what you type. Selecting a preset palette shifts the overall tone of the output; a custom palette, available on Ideogram's Plus plan or higher, takes "up to five colors", chosen with a picker or typed as hexadecimal values (docs.ideogram.ai, accessed September 2026).

That's the nuance worth holding onto: the same vendor whose prompting guide says hex codes aren't understood in written text does accept literal hex values through a different, structured parameter. Even there, Ideogram is honest about the ceiling: its own documentation describes the Color Palette setting as "best used as a creative direction tool rather than for precise color matching." The vendor's own dedicated color feature doesn't promise exact reproduction either.

What Color Words Actually Move the Output, If Hex Doesn't?

Named, visually grounded references, not raw codes and not vague mood words. Ideogram's prompting guide calls this pattern memory colors: "cherry red", "sky blue", "burnt orange", "mint green", real-world references that "evoke a mental image based on shared visual experience" rather than a number nobody can picture. Its own reference list groups dozens of these by base hue, then layers on modifiers for intensity, temperature and finish: pale, deep, muted, glossy, matte, iridescent.

Every other vendor's own worked examples lean on the same pattern without naming it. Google's Gemini image-generation guide, the one behind the Nano Banana Pro model covered in full in Prompting Gemini for Image Generation, fills a reusable logo template with "[color scheme]" as the placeholder, and one worked example resolves it as "The color scheme is black and white." A separate sticker example asks for "a vibrant color palette" in prose, and a style-transfer example describes "a dramatic palette of deep blues and bright yellows" (ai.google.dev, accessed September 2026). OpenAI's cookbook prompting guide does the same in its character-consistency example, listing "Same facial features, proportions, and color palette" as one of the things to hold steady across generations (developers.openai.com, accessed September 2026). Black Forest Labs' own guide lists tested color phrases directly: "muted beige and forest green tones", "deep blue and silver", "warm orange and pink sunset colors", "monochrome black and white", "desaturated pastel palette" (docs.bfl.ai, accessed September 2026).

None of these four vendors is quoting the others. Four independent sets of worked examples all converge on the same shape: a named hue plus a qualifier, not a coordinate.

How Do Intensity and Finish Modifiers Sharpen a Named Color?

By answering a question a bare hue never does: how bright, how warm, and how does light actually sit on the surface. Ideogram's own reference groups its modifiers into four categories rather than leaving them as one undifferentiated pile: intensity or brightness, "Pale, light, soft, faint, pastel, bright, deep, vivid, bold, dark, muted, rich, intense"; warmth and temperature, "Warm, cool, neutral, icy, frosted, fiery, dusky, earthy"; finish or surface quality, "Glossy, matte, metallic, shimmering, iridescent, pearlescent, velvet-like, translucent, silky, powdery"; and light and shadow modifiers, "Sunlit, shadowed, backlit, faded, dimmed, glowing, reflective, foggy, moonlit" (docs.ideogram.ai, accessed September 2026).

In table form (all items are from Ideogram's own list quoted above):

Its own worked examples combine one modifier from that list with one memory color rather than stacking several: "Powdery sky blue", "Rich cherry red", "Frosted mint green", "Dusky rose pink", "Shimmering sapphire blue", "Matte ivory white". Each pair answers a different question about the same hue, brightness from the first list, finish from the third, so rich cherry red and pale cherry red are legitimately different instructions rather than a stronger and weaker version of one color word. That's the same logic this site applies to lighting and material vocabulary elsewhere (see lighting vocabulary for AI image prompts and the anatomy of a great image prompt): a single qualifier changes what the model renders more than piling on vague, interchangeable praise adjectives ever does.

Does a Reference Image Beat Describing a Color in Words?

For transporting a whole palette at once, generally yes, because one image carries every relationship between colors, tint, saturation, how one hue sits next to another, where text has to spell each one out separately and can still miss how they interact. Midjourney's own Style Reference documentation, covered in depth in Midjourney --sref Explained, frames matching colors as one part of a broader package captured from an uploaded image, not a text field naming values: "It doesn't copy objects or people, just the overall style—like colors, medium, textures, or lighting—helping you achieve a consistent visual theme" (docs.midjourney.com, accessed September 2026).

That's also the vendor's own scoping worth keeping: a style reference carries color as one ingredient among several, not a dedicated, isolated color-matching tool. If the goal is one exact swatch and nothing else about the reference image, a written color word or, on FLUX.2, a hex code tied to a specific object, is the more targeted instruction. If the goal is an entire coherent look, a reference image does more work in one step than a paragraph of color adjectives.

What About Color-Grading Language Like Teal-and-Orange?

This is where cinematic vocabulary earns its keep, treated the same way this site treats lighting terms: real photographic and post-production convention, not confirmed prompt syntax any vendor defines. Black Forest Labs' own cinematic style reference lists "teal and orange color grading" alongside "anamorphic lens flare" and "film noir" as one row in a table of tested style phrases (docs.bfl.ai, accessed September 2026), which is the closest any of the four vendors checked here comes to documenting a named color-grading look rather than a single hue.

Grading language works as a mood-and-relationship instruction more than a hue instruction: "warm highlights, cool shadows" describes a relationship between two tonal ranges, which is a different kind of statement than naming one color. Combine grading language with a named palette rather than instead of one: "warm orange and pink sunset colors, teal and orange grading in the shadows" gives a model both the base hues and the tonal relationship, closer to what a colorist actually controls.

How Do You Prompt a Color Palette?

A palette prompt gives each color a role, so the model knows which one dominates. Name three colors and say where each goes:

[SUBJECT/SCENE], palette of three colors: [MEMORY COLOR A] as the dominant color covering most of the frame, [MEMORY COLOR B] as the secondary color on [WHERE], and [MEMORY COLOR C] as a small accent on [ONE ELEMENT]; no other hues

On a model that documents hex support, swap each memory color for a hex value tied to one named surface. Remember Ideogram's own caveat: its Color Palette setting is a creative-direction tool, not a precise matcher.

Copy-Paste Prompts for Color and Palette

Replace every [BRACKET] with your own text. Sixteen prompts: four named-color, four hex-code, four color-scheme and mood, three color-grading, and the one palette prompt above. Hex-code prompts below are written for models that document that support, FLUX.2 confirmed above, and should be adapted to a named color instead on a model that doesn't; the free FLUX.2 prompt generator and the free Ideogram 4.0 prompt generator are reasonable starting points for either path.

Named-color prompts (4)

[SUBJECT/OBJECT], [MEMORY COLOR, e.g. cherry red / sky blue / forest green], [MODIFIER: pale/deep/muted/glossy/matte], natural lighting consistent with that color's real-world appearance

[SCENE], color palette built around [MEMORY COLOR A] and [MEMORY COLOR B], the two colors appearing in [WHERE: background and foreground / subject and accent], nothing else competing for attention

[SUBJECT], desaturated pastel palette, [MEMORY COLOR] toned down toward gray rather than intensified, soft even lighting

[SUBJECT/SCENE], monochrome [COLOR] palette, every element rendered in shades of the same hue rather than a true black and white

Hex-code prompts, for models that document hex support (4)

[OBJECT] in color #[HEX], the hex value applied to [SPECIFIC SURFACE OR PART], rest of the scene lit to complement that color

[OBJECT], gradient from color #[HEX START] to color #[HEX END] across [DIRECTION: top to bottom / left to right], smooth transition with no banding

[SCENE] with a two-color custom palette: #[HEX A] for [ELEMENT A], #[HEX B] for [ELEMENT B], each hex value tied to a single named surface

Studio product shot, seamless gradient backdrop from #[HEX A] through #[HEX B], single hero product centered, soft specular highlight on the product itself

Color-scheme and mood prompts (4)

[SCENE], [color scheme placeholder, e.g. "warm orange and pink sunset colors" or "deep blue and silver"], lighting consistent with that scheme rather than fighting it

[SUBJECT], color scheme is [COLOR A] and [COLOR B] only, no other hues present, [STYLE: minimalist / bold / editorial]

[SCENE], vibrant color palette dominated by [COLOR], secondary accents in [COLOR B], everything else kept neutral

[SUBJECT/SCENE] in the artistic style of [REFERENCE], a dramatic palette of [COLOR A] and [COLOR B], preserving the original composition

Color-grading prompts (3)

[SCENE], warm highlights and cool shadows, teal and orange color grading, cinematic contrast without changing the base [SUBJECT] colors

[SCENE], film noir grading, high contrast, most color desaturated except [ONE ACCENT COLOR] left vivid

[SCENE], golden hour warmth in the highlights, [BASE PALETTE] holding steady in the midtones and shadows

A note on what Prompt Architects does here: it writes the prompt, not the image. Nothing here renders a picture; these are words to paste into whichever model does.