This guide takes you from a blank prompt box to an image you are actually happy with. You will learn what a text-to-image model really does, what every control in IMAGEKA changes, how to improve a prompt one deliberate step at a time, and where the tool's limits are so you do not waste generations fighting them.
- IMAGEKA runs several open-source image models and picks one for each request from your prompt and style preset. Whichever runs, it invents a new image from your words; it does not search for or copy an existing picture.
- Write a full sentence that names the subject, the setting, the light and the kind of image. The models read sentences well, so stacked keywords add little (anime prompts are the exception: short tags work well there).
- Change one thing per round and generate 2 or 4 images, so you can tell whether a change helped or you just got a lucky (or unlucky) random seed.
- Describe what you want, not just what you don't. The Negative Prompt field helps with photo, anime and painterly images, but the default general model largely ignores it.
- Your images are kept only in your browser. Download the ones you want to keep straight away.
What a text-to-image generator is (and isn't)
A text-to-image generator is a neural network trained on a very large collection of images paired with captions. From those pairs it has learned how words relate to visual patterns: what "overcast" does to shadows, what a "close-up" looks like, how a "watercolour" texture differs from an "oil painting". When you type a prompt, it starts from random noise and shapes that noise, over a few refinement steps, into an image that fits your words. If you want the mechanics, How AI image generators work explains them without the maths.
Three misconceptions are worth clearing up early, because they shape how you should write prompts:
- It is not a search engine. There is no library of finished images being looked up. Ask twice with the same words and you get two different, newly generated pictures.
- It does not understand your intent. It responds to what the words statistically tend to look like. "A cosy kitchen" works because thousands of captioned photos taught it what people call cosy. "A kitchen my grandmother would like" gives it almost nothing to work with.
- It is not a design tool. You cannot move an object ten pixels to the left, change only the shirt colour, or fix one hand. Each run is a fresh image. Precise edits happen afterwards, in an image editor.
A tour of the IMAGEKA interface
The generator on the home page has two panels: controls on the left, results on the right (on a phone, you switch between them with tabs). There is no account to create. Here is what each control does.
The prompt box
This is where you describe the image. It accepts plain English sentences, and that is how you should write. Under the box are sample prompt chips; clicking one fills the box, which is a quick way to see the kind of description that works. The model reads only a limited length of text (roughly the first couple of hundred words; the rest is cut off), and in practice two to four well-chosen sentences beat a long paragraph.
Enhance prompt
The Enhance prompt link adds a few descriptive phrases based on keywords it spots in your text. A prompt containing "portrait" gets something like "detailed eyes"; a landscape prompt gets lighting phrases such as "volumetric lighting". It is a simple rule-based helper, not an AI that rewrites your idea. It is handy for a two-word prompt, but read what it added and delete anything that conflicts with what you want.
The security code
Before each generation you type the five digits shown in the small security image. If the digits are hard to read, click the refresh icon for a new code. It keeps automated scripts from using up the free service. There is also a fair-use limit of 100 generations per day per IP address. On a shared connection, such as an office, school or café, everyone counts towards the same limit.
Settings
The Settings button opens four controls:
| Setting | Options | What it actually does |
|---|---|---|
| Art Style | Default plus 11 presets (Photorealistic, Digital Art, Oil Painting, Watercolor, Anime, Concept Art, Sketch, 3D Render, Minimalist, Vintage, Pixel Art) | Adds a fixed phrase to the end of your prompt and helps decide which model renders the image. Photorealistic is selected when the page loads. |
| Resolution | 512×512, 768×768, 1024×1024, 1024×768 (landscape), 768×1024 (portrait) | Sets the pixel size and shape. Shape changes composition as much as it changes size. |
| Images | 1, 2 or 4 | How many variations one run produces, each from a different random seed. |
| Negative Prompt | Free text | Things to steer away from. Works on the photo, anime and painterly models; largely ignored by the general model (see below). |
Style presets are less mysterious than they look. Choosing Watercolor, for example, adds "watercolor illustration, soft washes, paper texture, wet on wet technique" to whatever you typed. Because the preset is added after your words, it can clash with them. If your prompt already says "a pencil drawing" and the Photorealistic preset is still selected, the model gets two conflicting instructions and usually produces a muddled mix of both. The rule: if your prompt names a style, set the preset to Default. The art styles guide lists exactly what each preset adds and when it helps.
Presets do a second job: they steer which model IMAGEKA uses. Photorealistic or Vintage with a person, product or food subject goes to a photorealism model; Anime goes to an anime model; Oil Painting, Watercolor, Sketch and Pixel Art go to a painterly model; everything else uses a fast general model. Words in your prompt such as "photograph", "anime" or "watercolour" have the same effect. You can't pick a model directly, so the preset is your lever.
How much this field does depends on which model handles your request. The photorealism, anime and painterly models use it properly, and IMAGEKA already adds sensible defaults for them (bad hands, blurry, watermark and similar), so add only what is specific to your image, such as "glasses" or "text". The fast general model, used for posters, logos, landscapes and most other prompts, largely ignores it. For those, describe the absence in positive terms in the main prompt: "an empty, deserted beach at dawn" works far better than putting "people" in the negative field.
Results and the gallery
After you press Generate Image, the right-hand panel shows a status message and then your images. This usually takes from a few seconds to about 30 seconds, and a run of four images or a busy period takes longer. Each image has a download button. New images are added to a gallery that stays in place when you reload the page, until you clear it.
Your first prompt, worked step by step
Let's work through a realistic task: you need a header image for a blog post about baking bread at home. Keep the default settings for now (Photorealistic, 1024×1024, 1 image).
Round 1: the obvious attempt
bread
You will get a loaf of bread, probably on a plain surface and centred. The model has had to guess everything else: the setting, the light, the angle and the mood. The result is generic, and it changes a lot from run to run because nothing is pinned down.
A freshly baked sourdough loaf cooling on a wire rack on a wooden kitchen counter, flour dusted on the crust, soft morning light from a window on the left.
Subject, details, setting and light are all specified. The image now reads as "home baking" rather than as a stock photo of bread.
The second prompt follows a simple pattern: subject, then key details, then where it is, then the light. You don't need special keywords. Modern models read the relationships in a sentence ("on a wire rack", "from a window on the left") instead of treating your prompt as a bag of tags. The prompt-writing guide develops this structure further.
Round 3: change one thing, and only one
Suppose round 2 looks good but feels flat. Don't rewrite the whole prompt. Change a single element so you know what caused the difference. Here, change the light:
Set Images to 2 or 4 for this round. Each image uses a different random seed, and IMAGEKA picks a new seed every run. One image cannot tell you whether a change worked: you might simply have drawn a good or bad seed. Four images show you the range the prompt produces, and that range is what you are really judging.
Round 4: adjust the frame for where the image will go
A blog header is usually wide, so switch Resolution to 1024×768. A landscape frame leaves room around the subject, so add one phrase to use that space:
The model won't place things with pixel accuracy, but "on the right side" and "empty space on the left" are followed often enough to be worth saying. The colour and composition guide covers framing in more depth.
The loop, in short
- Write one clear sentence: subject, details, setting, light.
- Generate 2–4 images and look at what they have in common, not at the single best one.
- Name the biggest problem in one phrase ("too dark", "subject too small", "looks like a cartoon").
- Change the one part of the prompt or settings that addresses that problem.
- Repeat. Most images settle within four to six rounds. If they don't, the idea may be at the edge of what the model can do (see the next section).
Keep a copy of each prompt version in a notes file. The gallery stores images, not the prompts that made them, so a good prompt can be lost once you edit the box.
Realistic expectations: what the models do well and badly
IMAGEKA's models are chosen to be fast enough to run free at scale, and each is strong at its own kind of image: the general model at posters, text and concept scenes, the specialists at photos, anime and traditional media. They share some clear limits, though.
| Generally reliable | Often unreliable |
|---|---|
| Following descriptive sentences, including where things are ("a cat on the left, a dog on the right") | Exact counts beyond about three ("seven candles" often gives six or nine) |
| Photographic scenes: food, products, interiors, landscapes, portraits of invented people | Hands doing something specific, such as holding a pen or playing an instrument |
| Short text in quotes, such as a two- or three-word sign (quotes also route the request to the general model, which handles lettering best) | Long sentences, paragraphs or small print (letters get garbled) |
| Strong, simple styles: flat illustration, watercolour, pixel art | The same character looking identical across several images |
| Lighting and mood words | Negations ("no cars", "without a hat"), which often add the very thing mentioned |
Two further limits are features of IMAGEKA rather than of the model. The tool has no image-to-image, inpainting, upscaling or seed control. You cannot upload a photo to transform it, repaint part of an image, or regenerate an exact image later. If an image is 90% right, download it and fix the remaining 10% in an editor (crop, retouch, add your own text) rather than chasing perfection through regeneration. When things go wrong in a predictable way, such as extra fingers or warped faces, the troubleshooting guide has specific fixes.
For a sign or label, write it like this: a hand-painted wooden sign that reads "OPEN". One to three words in quotation marks have a good chance of being spelled correctly. Anything longer is usually better added afterwards in an editor, where you control the font.
Saving your images (read this before you generate a lot)
IMAGEKA does not keep your images on an account, because there is no account. The gallery lives in your browser's own local storage on the device you are using. That has practical consequences:
- Clearing browsing data deletes them. So can some privacy or "cleaner" tools, and so can the browser itself when local storage runs short of space.
- Private or incognito windows forget everything when you close them.
- They don't follow you. Images made on your laptop won't appear on your phone, or even in a different browser on the same laptop.
- Clear All really clears all. There is no recycle bin.
The habit that avoids disappointment: download every image you might want, as soon as you see it. Give the file a descriptive name, and paste the prompt into a text file saved next to it, so you can come back and make variations later.
Using it responsibly
A content filter blocks clearly harmful requests, such as sexual content involving minors or extreme violence, and a safety checker blurs outputs it judges to be explicit. Beyond what the filter catches, a few habits keep you out of trouble:
- Don't depict real people in situations they never were in, and don't pass off generated images of real events as photographs. Even when it is meant as a joke, a realistic fake can spread without its context.
- Avoid trademarks and copyrighted characters in anything you publish. The model may render a recognisable logo or character, but your right to use it commercially is a separate legal question.
- Label AI images where it matters: news-like content, product photos that customers rely on, and anywhere a platform's rules require it.
- Check the details before publishing. Look for an extra finger, a nonsense word in the background, or a product feature that doesn't exist.
All the models IMAGEKA uses are open-source, released under open licences that allow commercial use. Copyright in AI-generated images is less settled, though, and depends on where you live and how much human creative input went in. Before you put an image on packaging, ads or merchandise, read using AI images commercially.
Where to go next
Once the basic loop feels natural, the biggest quality jumps come from vocabulary. The lighting and camera terms guide teaches the words that most change a photographic image. The prompt library has tested prompts you can open directly in the generator and adapt. If you meet an unfamiliar term (CFG, LoRA, sampler), the glossary explains it and says whether it applies to IMAGEKA.
Frequently asked questions
Why do I get a different image every time with the same prompt?
Each generation starts from a different random noise pattern, set by a random number called the seed. IMAGEKA picks a new seed every run and doesn't let you set it, so identical prompts give different images. That's useful for exploring options. When you find one you like, download it, because you can't recreate it exactly.
Which model does IMAGEKA use? Can I choose a different one?
IMAGEKA uses several open-source models with different strengths: a fast general model (the default), a photorealism model, an anime model and a painterly model. It picks one automatically for each request based on your prompt and style preset, and if a specialist fails it retries with the general model. You can't choose a model by hand, but choosing a preset, or using words like "photograph", "anime" or "watercolour", effectively steers the choice.
Should I use the Enhance prompt button?
Use it when your prompt is very short and you want a quick starting point. It adds a few keyword-based phrases; it doesn't understand your idea. For anything specific, you will do better by writing the subject, setting and light yourself.
Do I need words like "8K, masterpiece, best quality"?
Not really. Those tags were a workaround for older Stable Diffusion 1.5 models, which were trained on scraped captions where such words correlated with nicer images. Current models respond much more to concrete description: what the subject is, what the light is doing, what kind of camera or medium the image looks like it came from.
Which resolution should I pick?
Pick the shape for where the image will be used: 1024×768 for headers and slides, 768×1024 for portrait posts and posters, 1024×1024 for square posts. Use 512×512 only for quick drafts, because it has noticeably less detail. The social media guide matches each size to common platforms.
Is there a limit on how much I can generate?
Yes. The fair-use limit is 100 generations per day per IP address, which is plenty for normal creative work. On a shared network, such as a school or office, everyone on that connection shares the same allowance, so make every round count by changing one thing at a time.
Where did my images go?
They were stored only in your browser. If you cleared browsing data, used a private window, or switched device or browser, they are no longer available, and there is no server copy to restore. Download images as soon as you make them.