Colour and Composition for AI Images

You cannot drag elements around in a text-to-image tool, but you can steer where things sit in the frame and which colours dominate, if you use words the model associates with layout and palette. This guide shows which composition and colour instructions IMAGEKA's models follow reliably, which it follows loosely, and how to plan images that leave room for a headline or logo.

Key takeaways

Say where the subject is ("on the left third of the frame", "small in the lower right corner") and what fills the rest ("the right two-thirds is open sky").
Negative space has to be described as something — a plain wall, an empty sky, a seamless backdrop — or the model fills it with detail.
The image size you pick changes the composition the model invents, not just the crop. Choose landscape, portrait or square before you write.
Named colours and small palettes ("navy, coral and cream") work well. Hex codes are not reliably followed, so fix exact brand colours in an editor.
Saturation and contrast words (muted, pastel, vivid, high-contrast) are some of the most dependable mood controls you have.

Subject placement

Left to itself, the model tends to put the main subject in the middle of the frame. That is fine for icons, profile pictures and symmetrical scenes, but it can feel static. You can move the subject with plain spatial language, and the models IMAGEKA uses understand it reasonably well. The fast general-purpose model, which handles most landscapes, cities and graphic layouts, is especially good at following placement written as a full sentence rather than a tag.

The rule of thirds divides the frame into a 3×3 grid. Placing the subject on a line or intersection, with open space in the direction it faces, usually reads as more dynamic than dead centre.

Phrases that move the subject

Centred: "centred composition", "the subject in the exact centre of the frame", "symmetrical".
Thirds: "positioned on the left third of the frame", "off-centre to the right". "Rule of thirds composition" works as a general hint but is less precise than naming a side.
Size in frame: "small in the frame", "tiny figure in a vast landscape", "fills the frame edge to edge".
Direction of gaze or movement: "looking towards the right side of the frame", "walking into the open space on the left". Subjects facing into empty space feel balanced; facing out of the frame feels tense.

Placement is followed often, not always. Generate four images and pick the one that landed; if none do, move the placement phrase earlier and describe what fills the other side.

Placement · left third

A lone red kayak on a still, misty lake, positioned on the left third of the frame and heading to the right. The rest of the frame is calm grey water fading into fog. Soft overcast light, muted colours. Minimal landscape photograph.

Negative space for text and logos

If an image will carry a headline, price or logo — an ad, a YouTube thumbnail, a blog header, a poster — plan the empty area in the prompt. "Leave space for text" helps a little, but the model does better when you describe what occupies that space, because every part of the image has to be something.

A layout that works for ads and headers: the subject on one third, a deliberately plain surface across the other two-thirds, and light falling towards the subject from the text side.

Good "empty" surfaces to name: a clear sky, a plain painted wall, a seamless paper backdrop, still water, fog, a smooth tabletop, a soft out-of-focus background. Avoid busy textures (brick, foliage, crowds) where text will go, because they fight legibility.

For thumbnails and banners, remember that IMAGEKA's widest size is 1024×768 (4:3). For a 16:9 frame, generate at 1024×768 with the subject placed well inside the frame, then crop the top and bottom. Keep anything important away from those edges. The social media guide lists safe crops for each platform.

Negative space · ad layout

A glass bottle of cold-pressed orange juice with condensation droplets, standing on the right third of a smooth pale stone surface. The left two-thirds of the frame is an empty, softly lit sand-coloured wall with nothing on it. Soft daylight from the left, a gentle shadow to the right of the bottle. Clean commercial product photograph.

Use 1024×768. The wall is described as a thing, so it stays empty instead of filling with props.

Leading lines, symmetry and framing

Leading lines

Lines in a scene — roads, rivers, fences, railway tracks, corridors, a row of lamp posts — pull the eye along them. Ask for the line and say where it leads: "a narrow cobbled lane leading from the bottom of the frame towards a lit doorway in the distance". The destination matters; a line that leads nowhere is just a road.

Symmetry

Symmetry feels calm, formal and deliberate. It suits architecture, interiors, tunnels and product hero shots. "Perfectly symmetrical composition, one-point perspective, centred" is a strong combination that the models follow well, especially for corridors and buildings. Reflections in water are a natural way to get symmetry in landscapes.

Framing within the frame

A doorway, window, archway, gap in foliage or the edges of a tunnel can surround the subject and focus attention. Write it as a relationship: "seen through an arched stone doorway", "framed by overhanging branches".

Framing · leading lines

A long, empty hotel corridor with patterned green carpet and rows of brass wall lights, perfectly symmetrical one-point perspective, the lines of the walls and ceiling leading to a single open door at the far end glowing with warm light. Wide-angle interior photograph, everything in focus.

Pick the Default preset so the Photorealistic preset's 85mm, shallow-focus wording does not soften the corridor.

Layers, depth and scale

Flat images usually have everything at one distance. Describing three layers — foreground, midground and background — gives depth, and the models handle these words well.

Foreground: something close to the camera, often partly out of focus: wildflowers, a railing, a cup on a table.
Midground: the subject.
Background: the setting, often softer or hazier: mountains, a skyline, a blurred café.

Scale cues tell the viewer how big things are. A tiny human figure next to a cliff makes the cliff enormous; a hand holding a miniature house makes the house small. If you want a sense of size, include an object of known scale and say how it compares: "a lone hiker, tiny against the base of the dam".

Depth · three layers

In the foreground, out-of-focus purple heather. In the midground, a stone shepherd's hut with smoke rising from its chimney. In the background, rolling hills fading into blue haze under a pale evening sky. Wide landscape photograph, gentle golden light from the left.

How image size changes composition

IMAGEKA offers square (512, 768 or 1024 pixels), landscape 1024×768 and portrait 768×1024. The model composes for the canvas it is given: the same prompt at landscape and portrait produces different arrangements, not crops of one image.

Size	Natural compositions	Good for
Square 1024×1024	Centred subjects, symmetry, close-ups, flat lays	Profile images, Instagram posts, icons, product hero shots
Landscape 1024×768	Horizons, subject on a third with space beside it, wide scenes	Blog headers, thumbnails, banners, slides
Portrait 768×1024	Full-length figures, tall buildings, layered top-to-bottom scenes	Posters, Pinterest, phone wallpapers, book covers

Two practical rules follow. First, pick the size before you write the prompt, and make the prompt agree with it: "a full-length figure" in a landscape frame tends to shrink the person or crop their feet. Second, use 1024-pixel sizes for final images. The 512×512 option is quicker for testing ideas but carries much less detail, and small faces and text suffer most. (When a request goes to one of the specialist models, IMAGEKA generates 512×512 at 768×768 instead, because those models look poor below that size, so a test can come back larger than you asked for.)

Choosing a colour palette

Colour words are among the most reliably followed parts of a prompt, especially when attached to objects or stated as a short palette. Two techniques work best:

Colour the objects: "a mustard-yellow armchair against a deep teal wall". This places the colours exactly.
Name a palette for the whole image: "colour palette of navy, coral and cream". This sets the overall feel and keeps a series consistent.

Specific colour names beat generic ones. "Terracotta", "sage green", "ochre", "burgundy", "dusty pink", "charcoal" and "cobalt blue" each narrow the range far more than "orange", "green" or "red".

A simplified screen-colour wheel. Complementary pairs give energy and contrast; analogous groups feel calm and unified; split-complementary keeps contrast but is easier to balance. On the traditional painter's wheel the pairs shift slightly (red is opposite green rather than cyan), but the principle is the same.

Colour schemes in words

Complementary: "teal and orange colour scheme", "red and green". High contrast and vivid; a common cinematic look is warm skin tones against cool blue shadows.
Analogous: "a palette of greens and blues", "warm palette of red, orange and amber". Harmonious and calm.
Limited palette: "only three colours: black, cream and rust red". Limiting the palette is one of the fastest ways to make an image look designed rather than generated, particularly for illustration and posters.
Monochrome: "monochromatic blue", "black and white". Strong for mood and for brand sets.
Accent colour: "a muted grey scene with a single bright yellow umbrella". The accent draws the eye exactly where you want it.

Limited palette · illustration

A flat vector illustration of a cyclist riding through a city park, simple geometric shapes, thick rounded outlines. Limited palette of only navy, coral and cream, no other colours. Lots of cream background around the cyclist.

Use the Default preset; the style is written into the prompt.

Saturation, contrast and mood

These words change the whole image at once and are followed consistently:

Saturation: vivid, saturated, bold for energy; muted, desaturated, faded for calm, nostalgia or seriousness; pastel for softness; earthy tones for natural, organic subjects; jewel tones (emerald, sapphire, ruby) for richness.
Contrast: high contrast, deep blacks for drama and graphic punch; low contrast, soft, hazy for dreaminess and gentle lifestyle imagery.
Brightness: bright and airy versus dark and moody. These overlap with high-key and low-key lighting, explained in the lighting and camera guide.

Colour also carries associations, though they vary by culture and context: blues often read as calm, trustworthy or cold; reds as energetic, urgent or passionate; greens as natural or fresh; warm yellows and oranges as friendly and optimistic; deep purples as luxurious. Use these as a starting point, then judge the actual image.

Mood · muted and nostalgic

A seaside promenade out of season, empty benches and a closed ice-cream kiosk with peeling paint, a single person walking a dog in the distance. Muted, faded colours of pale blue, sand and washed-out pink, low contrast, soft overcast light. Wide photograph.

Brand colours and hex codes

It is tempting to write "#1E6FD9" into a prompt. Image models do not reliably map hex codes to colours: they learned from captions written in words, and hex values are rare in that text. You may get roughly the right hue, a completely different one, or the code rendered as text in the image.

A more dependable approach:

Describe each brand colour in words, as precisely as you can: "a strong medium blue like a clear summer sky", "a warm coral pink".
Attach the colours to large, simple areas such as a wall, a backdrop or a piece of clothing, which are easy to recolour later.
Generate several options and pick the closest.
Adjust the exact colour in an image editor with a colour-replace or hue tool. IMAGEKA has no editing tools, so this step happens after download.

Logos are the same story: do not ask the model to draw your logo. Leave space and add the real file in a design tool. The small business guide has full recipes for branded images, and the commercial use guide covers why you should never prompt for other companies' logos.

Composition and colour checklist

Have you chosen the image size first, and does the prompt suit that shape?
Does the prompt say where the subject is and how big it is in the frame?
If text will be added, is the empty area described as a plain surface, on the opposite side from the subject?
Is there a foreground, midground and background, or a deliberate reason for a flat image?
Are colours attached to objects, or stated as a palette of two to four named colours?
Is there one saturation or contrast word that sets the mood?
Have you avoided hex codes and logos, planning to finish those in an editor?
Have you generated four images and judged which instructions landed consistently?

For how these layers fit into a complete prompt, see how to write AI image prompts.

Frequently asked questions

Why does my subject keep ending up in the middle?

Centring is the model's default. Name the position explicitly ("on the right third of the frame"), describe what fills the other side, and put the placement phrase early in the composition part of the prompt. Then generate several images, as placement is followed often but not every time.

Can I get exact brand colours?

Not reliably from the prompt alone. Describe the colours in words to get close, then correct them in an image editor. Hex codes are often ignored or rendered as stray text.

How do I make a 16:9 thumbnail?

Generate at 1024×768 with the subject and empty text area well inside the frame, then crop a strip from the top and bottom. Avoid placing faces or key details near those edges.

Does "rule of thirds" in a prompt actually work?

It is a mild hint. Naming a specific side ("on the left third") and saying what fills the rest of the frame is much more effective.