Midjourney Color Control: Hex Codes vs Named Palettes (Table)

Nine bottle renders, one palette, and three different greens: the first Marrow & Mint set came back with three sage bottles, two teal ones, and one the colour of a hospital corridor. All nine prompts contained the same three hex codes. The codes were not ignored so much as treated as suggestions, because a number in a prompt competes with every colour word, lighting term, and material reference around it.

Colour is the parameter that decides whether nine images read as one product family, and there are five ways to hold it. The table below ranks them by how much control they buy and what each one costs in prompt length and rework, followed by the exact fragment to paste for each and a drift test you can run on the whole set before anyone opens Photoshop. If you are still building the base prompt, the Midjourney prompt guide covers the parts that sit around the colour instruction.

How the Five Colour Methods Compare

MethodControl strengthPrompt costDrift over 9 framesBest for
Hex codes aloneLow to medium3 short tokensWide, 2 to 4 distinct huesA single hero image you will retouch anyway
Hex plus a named colour wordMedium6 to 10 tokensNarrow, usually 2 huesProduct shots with one dominant colour
Named palette list, no hexMedium12 to 20 tokensNarrow but shifts in value, not hueEditorial and lifestyle frames
Palette plus style referenceHighOne image URL and a weightSmall, 1 hue familyA line of products that must match
Style reference plus fixed seedVery highURL, weight, and a seed flagNear zero on reruns of one frameRegenerating one frame after a layout change

Read the last column first, because drift is the only thing the customer sees. A method that produces one beautiful bottle and eight near-misses is worse than a method that produces nine consistent ones, and the table is ordered by that logic rather than by how precise the instruction sounds.

What Hex Codes Actually Do in Midjourney

A hex code is read as text, not as a colour profile. The model has learned an association between the string and a region of colour space, and that association is loose. Put 2E5B4F in a prompt next to “frosted glass”, “morning light”, and “soft shadow” and the glass and the light will pull the result toward whatever green they imply. The code nudges; it does not clamp.

Two habits make hex codes behave better. Put them early, before material and lighting words, so they are not competing with a long tail of references. And pair each code with a plain-language colour name, because the model has far more training signal for “deep sage green” than for a six-character string. The pair costs six tokens and removes most of the ambiguity.

What hex codes cannot do is survive a subject change. A cap, a label, and a carton reflect light differently, and the same code will land at three different values across the three. That is not a bug in the prompt; it is the reason a reference image exists.

Three Levels of Colour Instruction

Level one, hex with a colour word. Use it when the frame is simple and you will retouch the final image anyway.

product photograph of a 50 ml frosted glass skincare bottle, deep sage green glass (hex 2E5B4F), bone white cap, standing on a pale limestone plinth, soft morning light from the left, shallow shadow, minimal background, 85 mm lens, commercial studio photography --ar 4:5 --style raw

Level two, a named palette. Better for lifestyle frames where exact values matter less than the relationship between colours. Name four colours and their roles, in order of area covered.

editorial still life for a skincare line, palette limited to four colours: muted sage 40 percent, bone white 30 percent, warm sand 20 percent, deep charcoal 10 percent. Two frosted bottles and a linen cloth on a plaster shelf, overcast daylight, no other colours in frame, matte finish, quiet composition --ar 4:5 --style raw --s 150

Level three, a reference image. The most reliable option for a nine-frame set, because the reference carries value and saturation in a way words cannot.

product photograph of a 50 ml frosted glass skincare bottle, sage green glass, bone white cap, limestone plinth, soft morning light, quiet minimal background, commercial studio photography --ar 4:5 --style raw --sref PASTE_REFERENCE_IMAGE_URL --sw 180

The weight is the dial. At 180 the palette dominates and composition loosens; at 60 the composition holds and colour drifts back toward the prompt’s own words. Marrow & Mint settled at 140 after testing three values on the same frame. The style parameter reference explains the surrounding flags, and our style reference guide goes deeper on how weights interact with aspect ratio.

The Nine-Frame Drift Test

Render all nine frames at the same aspect ratio and the same stylize value. Then compare them without touching them: place the nine thumbnails in one row and look for the frame that reads as a different product. In the Marrow & Mint set the outlier was always the dropper bottle, because a taller silhouette collects more light at the shoulder and the green shifts lighter.

Quantify it rather than argue about it. Sample the glass colour in the same relative spot on each bottle, convert to hue and saturation, and plot the nine values. A spread under 6 degrees of hue holds together in a grid. A spread over 12 degrees reads as two product lines.

You are reviewing nine product renders for Marrow & Mint. For each image I will give you the sampled RGB value from the centre of the bottle glass and from the cap. Convert each to HSL, report hue, saturation, and lightness, then flag any image whose hue is more than 8 degrees from the median. Suggest, for each flagged image, one change to the prompt that would pull it back: add a colour word, raise the style reference weight, or change the lighting term. Do not suggest retouching.

Run the test before the retoucher, not after. Nine images that need a colour grade cost more than nine re-renders, and a grade on a frosted surface rarely lands cleanly because the glass transmits the background as well as reflecting the light.

Cost and Failure Modes

Hex codes are free and unreliable. They cost three tokens and buy you a rough direction. The failure is quiet: the image looks plausible in isolation and wrong in a grid, and you only notice when the set is assembled.

A named palette costs 15 tokens and holds hue better than it holds value. The failure is a set that is consistent in colour but inconsistent in brightness, which reads as inconsistent under a shop light. Fix it by naming the light, not by adding more colour words.

A style reference costs a URL, a weight, and a licensing question. It also imports things you did not ask for: the reference’s contrast, its shadow softness, sometimes its composition. Marrow & Mint’s first reference image carried a hard shadow that appeared in all nine frames, and the fix was a second reference with the same palette and flatter light.

The most expensive option is none of these. Regenerating by hand until nine frames agree costs an afternoon per set and produces no reusable rule, so the tenth bottle starts the argument again. Pick a method, record the exact flags, and treat the palette as a spec rather than a mood.

Sources & Notes

Marrow & Mint, the nine bottles, the hex values, and the reference weight of 140 are an illustrative scenario, not a real brand or a measured benchmark. Midjourney’s handling of hex codes and style reference weights changes between model versions, so re-run the drift test after any version change rather than assuming last quarter’s settings still hold. The hue-spread thresholds are practical rules of thumb for product grids, not colour-science standards, and any image used in packaging should be checked against a physical proof.

声明:1.本站所有文章,如无特殊说明或标注,均为本站原创发布。本网站中【文本生成/文生图】使用了DeeSseek人工智能生成内容技术, 算法来源说明:本网站未自行开发、训练或部署深度合成算法,所使用的人工智能算法来自:DeepSeek、Chatgpt、Google Gemini。 2.为了网站加载速度优化,本站预览图片进行压缩处理,所以有些不是太清晰,并不影响AI提示词生图、AI设计等内容。3.任何个人或组织,在未征得本站同意时,禁止复制、盗用、采集、发布本站内容到任何网站、书籍等各类媒体平台。如若本站内容侵犯了原著者的合法权益,可联系我们进行处理。--GUUNN.COM