Prompt Weights and Order: Does Word Position Matter?

Updated: 2026-11-03 · 6 min read

Swap two phrases in a prompt and the picture can change completely. Put "watercolour" first and you get a painting; put it last after forty words of detail and you may get a photo with a faint wash. Word order isn't magic, but it isn't random either.

Why position can matter

Image generators convert your prompt into numbers using a text encoder, then use those numbers to guide image creation. Exactly how word position affects the result depends on the model and its text encoder, and it varies between tools. In practice, many users and tests find that:

• Earlier words often carry more weight, especially in keyword-style prompts.
• Very long prompts may be truncated or diluted in some models, so later details get less attention. Some encoders have a token limit, and words beyond it may be ignored or handled differently by the interface.
• Conversational models that understand full sentences are less sensitive to order but still benefit from clarity.

Because behaviour differs, treat these as tendencies to test, not laws.

The practical order that works almost everywhere

1. Medium / image type: "35mm photograph", "watercolour painting", "flat vector icon".
2. Subject: who or what, with key traits.
3. Action and setting.
4. Lighting.
5. Camera and composition.
6. Mood, palette and finishing details.

This order mirrors how a person would describe an image and puts the most structural decisions first. It's the same idea behind our beginner prompt formula.

A simple experiment you can run

Try these two prompts with the same seed (if your tool allows) or several generations each:

A: watercolour painting of a red fox in a snowy forest, soft morning light

B: a red fox in a snowy forest, soft morning light, detailed fur, tall pine trees, falling snow, small tracks in the snow, rabbits in the distance, watercolour

In many generators, A looks clearly like watercolour while B leans toward a detailed illustration or photo with a watercolour hint. The medium moved from first to last and was buried under details.

Run your own tests:

• Swap the order of two adjectives, such as "golden hour, foggy" versus "foggy, golden hour".
• Move the subject from the end to the start.
• Remove half of the details and see what changes.

Explicit weights: Midjourney

Midjourney supports multi-prompts using a double colon ::. It separates concepts so they're considered independently, and you can add a number after the colons to set relative importance:

hot dog::1.5 street food stand::1

space::2 ship::1

Negative weights can reduce a concept (Midjourney's --no parameter works similarly, equivalent to a weight of -0.5 according to its documentation). Syntax and behaviour can change between model versions, so check the official docs. For a full breakdown, see Midjourney prompt structure.

Explicit weights: Stable Diffusion interfaces

Popular Stable Diffusion front ends support emphasis syntax:

• (word) increases attention slightly.
• (word:1.3) sets a specific weight; values above 1 increase emphasis, below 1 reduce it.
• [word] decreases attention in some interfaces.

Example:

portrait of an old sailor, (weathered skin:1.3), (blue knit cap:1.1), soft overcast light, 85mm, (background:0.7)

Exact syntax depends on the interface (AUTOMATIC1111-style and ComfyUI handle weights in their own ways), so check your tool's documentation. Weights above about 1.5 often cause distortions; small adjustments usually work better.

Negative prompts have their own weighting too; see negative prompts explained.

Conversational tools: clarity beats order

ChatGPT and Gemini read full sentences and follow instructions, so you can say "The most important thing is that the fox is red and clearly visible in the centre." That's an explicit priority written in plain language. Order still helps readability, but these tools don't use :: or (word:1.3) syntax. See our ChatGPT and Gemini guides.

Repetition as a weight

Repeating a concept can increase its influence in some models: "a cosy cabin, warm cosy interior, cosy lighting". It's a blunt tool, and it can also make images over-stylised or repetitive. Prefer explicit weights where supported, or move the concept earlier.

Common order mistakes

• Starting with quality buzzwords: "masterpiece, best quality, 8k, ultra detailed, a dog..." pushes the actual subject later. Many newer models don't need these words at all.
• Style at the very end of a long prompt: move it to the front.
• Conflicting concepts close together: "minimalist, cluttered market" confuses the model whatever the order.
• Important details hidden in the middle of a list: if the red umbrella matters, mention it early and clearly.
• Overly long prompts: after a point, extra words add noise. Cut anything that doesn't change the image.

Prompt structure for different goals

When style matters most: medium → subject → setting → light.

ukiyo-e woodblock print of a lighthouse on a rocky coast during a storm, large curling waves, muted indigo and cream palette

When the subject must be exact: subject → traits → action → medium.

a white ceramic teapot with a bamboo handle, slightly turned, on a sage-green background, studio product photo, soft light

When mood matters most: mood/light → subject → setting.

eerie blue moonlight, an abandoned carousel in a foggy park, cinematic wide shot

How to know what's working

Keep a simple log: prompt, tool, settings and a one-line note on the result. After a dozen tests you'll see your tool's preferences. That habit, more than any trick, is what makes prompt writers consistent.

Start with well-ordered prompts

Ready-made prompts are useful learning material: you can study how they're ordered and what each phrase does. The AI Image Prompts Android app gives you a categorised library with search, preview images, one-tap copy and favourites, and it's updated regularly. It's a prompt library rather than a generator; paste prompts into your AI tool, then experiment by moving the medium or subject to see how order changes your results.

A worked example: fixing a prompt by reordering

Here's a real-world style problem. The prompt below was meant to create a children's book illustration, but kept producing a photo:

a little girl in a yellow raincoat jumping in puddles on a city street, rainy day, reflections, umbrellas, people walking, shop windows, detailed, children's book illustration

Three changes fixed it, without adding a single new idea:

1. Medium first: "Children's book illustration of a little girl in a yellow raincoat jumping in puddles…"
2. Fewer competing details: "people walking" and "shop windows" were removed, leaving the girl, puddles and a few umbrellas.
3. Style reinforced with technique words: "soft gouache textures, rounded shapes" replaced "detailed".

The new prompt is shorter, the subject is unmistakable and the medium leads. That is the whole lesson of word order in one example: decide what matters most, put it first, and remove anything that competes with it.

FAQ

Does word order matter in AI image prompts?

Often, yes. In many generators earlier words have more influence, so put the medium and subject first. The effect varies between models, so test.

How do I add weight to a word in Midjourney?

Use multi-prompt syntax with a double colon and a number, for example space::2 ship::1. Check the docs for your model version.

What does (word:1.3) mean in Stable Diffusion?

It increases the emphasis on that word to 1.3 times normal in interfaces that support the syntax. Values below 1 reduce emphasis.

Do ChatGPT and Gemini use prompt weights?

No special syntax. State priorities in plain language, such as "the most important element is…".

Should I start prompts with "masterpiece, best quality"?

It's usually unnecessary with modern models and pushes the subject later. Start with the medium and subject instead.