This guide gives you a repeatable way to write prompts for IMAGEKA: an order to put your ideas in, the kind of language its models follow best, how your words and preset decide which model draws the image, and a method for improving a prompt in a few runs instead of a few dozen. Every rule is shown with a worked example you can open in the generator.
- Write a short paragraph in plain sentences. IMAGEKA's fast general model reads relationships ("a cat on top of a fridge") far better than older image models did. The anime model is the exception that also likes short comma-separated tags.
- IMAGEKA picks one of several models to suit your prompt. Your preset and words steer the choice: Photorealistic plus a person, product or food goes to the photorealism model; Anime goes to the anime model; Oil Painting, Watercolor, Sketch and Pixel Art go to the painterly model; quoted text, logos and posters go to the general model.
- Put the subject first, then details, setting, composition, light and finally medium or style. Earlier words tend to carry more weight.
- Concrete nouns and visible details beat praise words. "Beautiful" and "masterpiece" add almost nothing; "chipped enamel mug" adds a lot.
- Describe what you want to see. The negative prompt works on the photo, anime and painterly models but is largely ignored by the general model, so "an empty beach" is more reliable than "beach" plus a negative of "people".
- Change one thing per run, and keep the wording that works as reusable "anchor" phrases.
How IMAGEKA reads your prompt
IMAGEKA does not send every prompt to the same model. It runs several open-source image models, each with a different strength, and picks one per request by looking at your words and the style preset you chose. You cannot pick the model by hand, but you can steer the choice, and it helps to know the four roles:
| Model role | Chosen when | Good at |
|---|---|---|
| Fast general model (the default) | Text in quotes, logos, posters, icons, stickers, diagrams; anything no specialist claims | Following long natural sentences, short lettering, graphic design, concept art, landscapes, cities |
| Photorealism model | A person, portrait, product or food together with the Photorealistic or Vintage preset, or with photo words such as "photograph", "portrait", "85mm" or "close-up" | Skin texture, believable light, product and food shots |
| Anime model | The Anime preset, or words such as anime, manga, chibi or cel-shaded | Anime and manga characters and scenes; responds well to comma-separated tags |
| Painterly model | Oil Painting, Watercolor, Sketch or Pixel Art presets, or words such as watercolour, charcoal, gouache, impasto, linocut or pixel art | Traditional media and retro pixel looks |
The checks run in that order of priority: quoted text and design words such as "poster" or "logo" send a prompt to the general model even if the Anime or Photorealistic preset is selected, because that model is the most reliable at lettering and layout. If a specialist model fails for any reason, IMAGEKA retries with the general model automatically.
Two facts about the models shape how you should write:
- The general model understands grammar. Its text encoder is built on a large language model, so word order and how parts of a sentence relate matter. That is why "a red cube on a blue sphere" usually gives you the red thing on top rather than a random mix of red and blue shapes. Older models relied on simpler encoders, which is why people used to write strings of comma-separated keywords. The specialists understand sentences too, but they give most weight to the first part of a prompt, so front-loading matters even more for them.
- There is a length limit. Every model reads only a limited amount of text, and words past that point are weighted weakly or cut off, so a sprawling prompt can lose its ending. Most good prompts are 30–80 words. The how it works guide explains text encoders in more detail.
A prompt structure that works
You do not need a formula, but you do need an order. When every prompt follows the same sequence, it is easy to see what is missing and to change one part at a time. This is the order we use:
- Subject — who or what the image is about, with a number if it matters: "two elderly men", "a single paper lantern".
- Details — what the subject looks like and is doing: clothing, materials, expression, pose, action.
- Setting — where it is and what surrounds it: "in a cluttered bicycle repair shop", "on a wet slate rooftop".
- Composition — how it is framed: shot size, angle, where the subject sits in the frame. See colour and composition.
- Light — the direction, quality and colour of light, and the time of day. See lighting and camera terms.
- Medium and style — photograph, watercolour, linocut, 3D render; plus colour palette and mood.
Subject comes first because it is the one thing the image cannot be without. Medium comes last because it applies to everything before it, and because IMAGEKA's style presets append their text at the very end of your prompt, so your own style words sit next to the preset's.
Why sentences beat tag lists
A tag list gives the model a bag of concepts with no relationships between them. It has to guess which adjective belongs to which noun, and it often guesses by blending. Sentences remove the guessing: grammar tells the model that "blue" belongs to "door" and "yellow" belongs to "bicycle".
woman, red coat, blue umbrella, rain, street, night, neon, cinematic, 8k, masterpiece, best quality, ultra detailed
Colours can swap between coat and umbrella, "night" and "neon" have no location, and the last five tags are habits carried over from older models. They use up tokens without describing anything you could point at.
A woman in a long red wool coat holds a transparent blue umbrella as she crosses a rain-soaked street at night. Pink and green neon signs reflect in the puddles around her feet. Shot from across the street at eye level.
Every colour is attached to an object, the neon has a job (reflections), and the viewpoint is stated. Nothing is there just for luck.
Short keyword phrases are still fine where they read naturally, especially at the end for style: "watercolour, soft washes, muted palette". The point is not to avoid commas; it is to make sure every object has its own description attached to it.
The anime model is a partial exception. It learned from images labelled with short tags, so a compact tag list such as "1girl, silver hair, school uniform, cherry blossoms, wind, looking back" works well there. A sentence works too, and mixing both (tags for the character, a sentence for the scene) is often the most reliable. Use the Anime preset so the prompt is routed to that model.
Word order and specificity
Front-load what matters most
IMAGEKA's models are much better than older ones at honouring a whole prompt, but when a prompt is crowded, details near the end are the first to be dropped or blended. If the image must contain a particular object, mention it early. If something is mere atmosphere, it can go later.
Use nouns you could photograph
The model learned from images paired with captions, and captions describe visible things. "A cosy kitchen" is a feeling; "a small kitchen with a cast-iron stove, copper pans on hooks and a steaming kettle" is a set of objects that together produce cosiness. Replace abstract adjectives with the visible evidence for them.
Be specific where it counts, vague where it does not
Over-specifying everything can backfire: ten precise demands compete, and some will lose. Decide which two or three details are non-negotiable and describe those carefully. Leave the rest loose so the model can fill it plausibly.
State counts and positions plainly
"Three ceramic cups in a row, the middle one tipped over" is followed more often than "some cups, one tipped". Small numbers (one to four) are reliable; larger counts are not, so "a crowd" is safer than "seventeen people". Spatial words such as left, right, behind, in the foreground and on top of are understood, though not perfectly — check the result.
Five worked rewrites
Each pair below starts from a prompt we see often and rewrites it using the structure above. The reasoning matters more than the exact words.
1. Product shot
candle product photo, luxury, aesthetic, high quality
"Luxury" and "aesthetic" are judgements, not instructions. The model has to guess the candle's shape, the surface, the background and the light, so results vary wildly and look generic.
A soy candle in a frosted amber glass jar with a plain kraft paper label, lit, sitting on a pale travertine slab. Behind it, a softly blurred linen curtain. Soft window light from the left, gentle shadow falling to the right. Clean, warm commercial product photograph with plenty of empty space above the candle.
Materials (frosted amber glass, travertine, linen) do the "luxury" work. A blank label avoids garbled lettering. Light direction and empty space make it usable as an ad with text added later.
2. Portrait
portrait of a beautiful old man, detailed face, realistic
"Beautiful" pulls towards a polished, stock-photo look. "Detailed face" says nothing about which details. There is no setting, light or framing.
Close-up portrait of a man in his seventies with a short white beard, deep laugh lines and a faded flat cap, looking slightly past the camera with a small smile. Dark wooden pub interior behind him, out of focus. Warm light from a window on his left side, the other half of his face in soft shadow.
The age, beard, wrinkles and cap give the face character. The off-camera gaze and small smile set the mood. One-sided window light adds the depth that "realistic" was hoping for.
3. Landscape
epic mountain landscape, stunning, 4k, trending
"Epic", "stunning" and "trending" are popularity words. With no time of day, weather or foreground, you get a flat, postcard-like range of peaks.
A narrow glacial lake in a steep valley, jagged granite peaks with fresh snow on both sides. Low mist drifting across the water. A lone wooden rowing boat pulled up on the pebbles in the foreground. Blue hour just after sunset, the last pink light on the highest summit. Wide landscape photograph, the horizon on the upper third.
Weather (mist) and time (blue hour, last light on the summit) create the drama. The rowing boat gives scale and a foreground, which makes the scene feel deep rather than flat.
4. Illustration
cute fox illustration for kids book
"Cute" and "kids book" leave the medium open, so you may get anything from 3D cartoon to glossy digital painting. Nothing tells the model what the fox is doing.
A small red fox wearing a yellow knitted scarf, sitting on a tree stump and reading a tiny book, in an autumn forest with falling leaves. Children's picture-book illustration in gouache, flat matte colours, visible brush texture, soft rounded shapes, warm muted palette of rust, mustard and sage green.
An action gives the image a story. Naming the medium (gouache), its texture and a three-colour palette pins down the look, so a series of images will match.
5. Poster with text
poster for a jazz night at a bar called The Blue Room on friday 8pm with live music and drinks
The model tries to render every word as text, invents spellings, and has no idea what the poster should look like.
A vintage-style jazz concert poster. A silhouette of a double bass player against a deep navy background with a single cream spotlight. At the top, large bold cream lettering reads "BLUE ROOM". Screen-printed look, two colours, slight paper grain.
Only two short words are requested as text, in quotes, with a placement and colour. Date and time are left out; add them in an editor where they will be spelled correctly.
Writing prompts with text in the image
Lettering is the general model's speciality, and putting words in quotes is also what routes your prompt to it. It renders short text far more reliably than older models, but it still misspells, especially with long or unusual words. These habits raise your hit rate:
- Put the exact words in double quotes and introduce them with a verb: a sign that reads "OPEN LATE". The quotes also make sure the general model, the one best at lettering, handles the image.
- Keep it short. One to three words works best. A full sentence will usually contain at least one error.
- Say where the text is and what it is on: "painted on the shop window", "embroidered on the cap", "at the top of the poster".
- Describe the lettering: "bold sans-serif capitals", "hand-painted script", "neon tubing". It helps the model choose a plausible form.
- Generate four at once and pick the one that is spelled correctly. Spelling is partly luck from the random seed.
- Add anything longer afterwards. For dates, prices, addresses and body copy, leave empty space in the image and type the text in a design tool. It will be correct and editable.
Keeping a character or look consistent
IMAGEKA picks a new random seed for every run and has no image-to-image or reference upload, so you cannot lock a character the way some paid tools allow. You can get surprisingly close with anchor phrases: a fixed block of description that you paste word for word into every prompt.
A good anchor describes features that are distinctive and easy to see, in the same order every time. Distinctive matters because generic features ("brown hair, blue eyes") are shared by countless faces, so each run picks a different one. Unusual combinations narrow the options.
Tips for anchors:
- Keep the anchor at the start of the prompt so it is never truncated or crowded out.
- Clothing and accessories carry more consistency than facial structure. A signature jacket is easier to repeat than a particular nose.
- Fix the style sentence too. The same character in a different medium will look like a different person.
- Expect a family resemblance, not an identical person. Generate several and choose the closest match.
The same technique keeps a brand look consistent across a set of images: write a style anchor ("flat vector illustration, thick rounded outlines, palette of navy, coral and cream, plenty of white background") and end every prompt with it. The social media guide shows how to use this for a feed.
How style presets interact with your prompt
A style preset in IMAGEKA is not a filter. It appends a fixed phrase to the end of your prompt before it is sent, and it also helps decide which model draws the image (see how IMAGEKA reads your prompt). Photorealistic is selected by default and adds "photo taken on Sony A7R, 85mm lens, shallow depth of field, photorealistic, ultra sharp". The full list is in the art styles guide.
That has three practical consequences:
- If your prompt names a style, switch the preset to Default. "A linocut print of a heron" plus the Photorealistic text asks for two incompatible things, and you will often get a photo of a print, or a muddy compromise.
- Watch for hidden camera instructions. The Photorealistic preset asks for an 85mm lens and shallow depth of field. If you want a wide-angle interior with everything sharp, those words work against you. Choose Default and write your own camera terms.
- Presets steer the model choice. If you want the photorealism model for a portrait, keep Photorealistic (or Vintage) selected; for anime, choose Anime; for paint, pencil or pixel looks, choose the matching preset. With Default you can get the same effect by writing words such as "photograph", "anime" or "watercolour" into the prompt.
- Presets are a good starting point for short prompts. If you only want to describe the subject, a preset supplies a reasonable medium. Once your prompt grows its own style sentence, the preset becomes redundant.
The Enhance prompt button works in a similar spirit: it adds a few descriptive phrases based on keywords it spots (portraits get "detailed eyes", for example). It is a quick helper, not a rewriter, so read what it added and delete anything that conflicts with your intent.
Say what you want, not what you don't
What the Negative Prompt field does depends on which model handles your image. The photo, anime and painterly models use it properly, and IMAGEKA already adds sensible default negatives for them (bad hands, blurry, watermark and similar), so you do not need to type those yourself. Add negatives there only for specific things that keep appearing, such as "hat" or "second person". The fast general model, which handles text, posters, landscapes and most other prompts, largely ignores the field.
Because you do not always know which model will be chosen, positive phrasing is still the habit that works everywhere. Also, mentioning an unwanted thing in the main prompt can add it: "a street with no cars" puts the concept cars into the prompt. Describe the state you want instead.
| Instead of | Write |
|---|---|
| a beach, no people | a deserted beach at dawn, untouched sand, empty to the horizon |
| portrait, not blurry | portrait, sharp focus on the eyes |
| a room without clutter | a minimal room with bare white walls and a single chair |
| no text, no watermark | a plain unmarked background (and avoid words like "poster" or "label") |
| not cartoonish | a documentary photograph with natural skin texture |
If text keeps appearing, look for words that imply it — sign, book cover, packaging, magazine — and rephrase them. For anatomy problems such as hands, see the troubleshooting guide.
An iteration method that saves time
Because the seed changes every run, two runs of the same prompt differ. That randomness is useful for exploring, but it makes it hard to tell whether a change in wording helped. A simple method keeps you from chasing noise:
- Write the first draft using the six-part structure and generate four images. Four images show you what the prompt reliably produces versus what was luck.
- Judge the set, not the best image. If three of four miss the same thing, the prompt is the problem. If only one misses, it was the seed.
- Change one layer at a time. Fix the subject first, then composition, then light, then style. Changing everything at once tells you nothing.
- Strengthen rather than repeat. If a detail is ignored, move it earlier and describe it more concretely. Repeating the same word three times rarely helps.
- Keep a text file of winners. Copy prompts that worked, along with notes such as "Default preset, 1024×768". IMAGEKA stores images only in your browser, so download the images you want to keep.
More prompts to start from
These cover common use cases. Each shows the structure at work; swap in your own subject. The prompt library has many more, sorted by category.
Prompt checklist
- Is the subject in the first few words, with a number if quantity matters?
- Does every colour and material belong to a named object?
- Have you replaced judgement words (beautiful, epic, stunning) with visible details?
- Is there a setting, a framing and a light source?
- Is the medium stated once, and is the preset set to Default if you named a style yourself?
- Is every unwanted element rephrased as a positive description?
- Is any requested text short, in quotes and placed?
- Is the prompt under about 150 words, with the essentials near the start?
- Have you chosen an image size that suits the composition (landscape, portrait or square)?
Frequently asked questions
Do quality words like "8K" or "masterpiece" help?
Very little. They were popular with early open models because training captions associated them with higher-rated images. IMAGEKA's models already aim for clean, detailed output, and the specialists get suitable quality terms added automatically. Spend those words on describing the subject, light or material instead.
How long should a prompt be?
Most good prompts are 30–80 words: long enough to cover subject, setting, framing, light and style, short enough to stay focused. Beyond roughly 150 words you risk the end being cut off, and extra details start competing with each other.
Should I still use the Negative Prompt field?
Yes, for specific unwanted things. It works on the photo, anime and painterly models (which already get default negatives such as bad hands, blurry and watermark), but the fast general model largely ignores it. So use it as a backup, and rewrite the main prompt to describe what you want first: that works whichever model is chosen.
Why does the same prompt give different results each time?
Each run starts from a different random seed, the noise pattern the image grows from. IMAGEKA does not offer seed control, so generate several images per run and keep the best, or use anchor phrases to keep important features stable.
Can I write prompts in languages other than English?
The models' text encoders were trained mostly on English, so English prompts are the most reliable. Other major languages often work for simple descriptions, but detail and accuracy usually drop. Text you want rendered in the image can be in another language, though accented letters and non-Latin scripts are more error-prone.
Can I fix one part of an image without regenerating it?
Not in IMAGEKA; there is no inpainting or editing. Either adjust the prompt and regenerate, or download the image and make the change in an image editor.