Ask a text-to-image model for three red apples on a wooden table and you will usually get a handsome table, lovely light, and a number of apples that may or may not be three. That one request holds most of what there is to know about prompting for images: the words steer the feel of a picture well and its facts badly.
The AI-Powered Image Generation tool on my projects page is built for exactly this kind of trial and error. It is a Gradio interface over Python: type a description, see the image, change the words, go again. Spend a while in that loop and you learn where the words have a firm grip and where they barely hold on.
Where the words have a firm grip
- Subject. "A lighthouse on a cliff" gets you a lighthouse on a cliff. The main thing in the picture, and framing words like close-up, wide shot or seen from above, are followed reliably.
- Style. Watercolour, pencil sketch, flat vector illustration, an old film photograph. Style words are some of the strongest in any prompt, because models learn from enormous numbers of images described in exactly those terms.
- Lighting and mood. Golden hour, overcast, neon at night, a single candle. Light changes a picture more than almost any other word, and models respond to it consistently.
- Colour, in broad strokes. "Muted earth tones" or "black and white" works. "This exact shade of blue" does not.
What these have in common is that they describe the whole image, or the one dominant thing in it. The model is building an overall impression, and impressions are what it is best at.
Where the words barely hold
- Exact counts. Two of something often works. Seven usually does not. Nothing in the process counts; the model draws something that looks like "several apples", and the number falls out of that.
- Text in the image. A shop sign, a label, a name on a cake. Many models produce shapes that look like letters from across the room and spell nothing up close. Some newer models manage a short word or two far better than older ones, but exact spelling across a sentence is still not something to rely on.
- The same character twice. Describe a woman with short grey hair and a green coat, generate again, and you get a different woman with short grey hair and a green coat. Words alone do not pin down a face. That takes extra machinery, such as reference images or a small model trained on that one character.
- Who has which attribute. "A red cube on a blue sphere" sometimes comes back as a blue cube, or a cube beside the sphere. Colours and positions attach loosely to the nouns near them, and the attachment slips.
A prompt is a firm request about mood and a loose request about facts. Plan the work around that and you stop fighting the tool.
What helps
None of these fixes the weak spots completely. Together they turn hoping into working.
Negative prompts
Many diffusion models accept a second prompt listing what you do not want: blurry, watermark, text, extra fingers. It is the most direct way to remove a flaw that keeps coming back. Not every model supports one, and writing "no text" in the main prompt tends to backfire, because mentioning text puts the idea of text into the picture.
Seeds for repeatability
Every image starts from random noise, and the seed decides which noise. With the same model, settings, prompt and seed, you generally get the same image back. That turns prompting into an experiment: fix the seed, change one word, and you see exactly what that word did. Keep the prompt and change the seed, and you see the range of pictures the prompt allows. Where a tool exposes the seed, note down the ones that worked.
Generate several, then choose
Because counts and details are unreliable, never judge a prompt by one image. Generate four and keep the one that got the apples right. That is not a workaround; it is how the method is meant to be used. For someone hunting for an idea, four rough options in quick succession beat one careful attempt.
Put first what matters most
Long prompts dilute themselves. Many models give more weight to what comes early, and some silently ignore everything past a fixed length. Lead with the subject, then style and light, and stop before the prompt becomes a paragraph.
Finish outside the generator
When a picture needs exact words or an exact count, add them afterwards in an image editor. Typing a word once is quicker than asking a model to spell it twenty times.
A direction, not a specification
Treat the first image as a sketch from a talented collaborator who did not quite listen. Keep what works, change the words for what does not, and go again. If you want to try the three apples yourself, the tool is on my projects page.
