Making an image

Hi, I'm Mel. Suppose you need a photograph that nobody took: a product concept, an imaginary place, or tonight's burger before anyone cooked it. An image generator can make it, but the first result depends heavily on what you asked for. I tested that with one meal, two prompts and a third instruction that exposed the difference between generating another picture and changing the one I already had.

I started with one sentence: create a landscape image of a burger and fries in a restaurant. The result is attractive. It is also the generator's photograph, not mine. It chose the camera position, the plate, the drink, the restaurant, the styling and even some blurred lettering on the wall. I named the subject and left every photographic decision open.

So I asked again, this time as a photographic brief: an editorial food photograph, a smashed cheeseburger on a dark cafe table, wide landscape, burger left of centre, eye-level close crop, shallow depth of field and warm side light. Then the useful exclusions: no people, hands, text, logos, cutlery or extra food. The subject stayed the same. The level of creative control changed completely.

A useful prompt does not need to be long. It needs to settle the decisions you care about. Name the subject. Set the composition and viewpoint. Describe the light. Add the visual character or mood. Then state the constraints that protect the result. Camera language can help steer the look, but it is direction rather than a physical guarantee. If a detail affects whether you can use the image, say it plainly.

Next I changed only one phrase in the written brief: warm side light became cool morning light. But I generated the prompt again in a fresh chat. The model did not revise the first photograph. It made another one. The light changed, but so did the burger, the fries, the arrangement and the serving plate. The prompt had narrowed the choices, but it had not frozen the previous image. A fresh generation is another interpretation, even when most of the words are identical.

So I returned to the original image and changed the job. I said: change only the lighting to cool morning light from the left. Then I named what had to remain: the burger, fries bowl, glass, camera angle, framing, object positions, ingredients, background and crop. This time the scene held together and the light moved. Generate when you want another interpretation. Edit when the original image needs to survive.

I also tested two warnings that used to be reliable. I asked for exactly seven dumplings, and the image contained seven. Then I generated the same four-word cafe sign three times. All three were correct. That does not prove every count or every word will be right. It proves the old rule that image generators cannot handle text or numbers is now too crude. Ask clearly, then inspect what came back.

The checking step stays because the image can look finished before it is dependable. Count the objects. Read every word. Check hands, reflections, packaging and any product detail that has to match reality. If the picture is evidence of a real person, place or product, use a real source image rather than asking a generator to invent accuracy. The polished surface is not the proof.

There are three sensible routes. Use a general assistant such as ChatGPT and Gemini when you want to describe and refine the image through conversation. Use Midjourney when sustained art direction and a repeatable visual style are the job. Use Canva's generator when the image is going straight into a design you are already building. None is the winner for every job. Choose the route that matches what happens after the image appears.

The method is simple. Describe the photograph, not only the subject. Generate again when you want another possibility. Edit the existing image when continuity matters. And check the result before it leaves your screen. A good prompt gives the generator direction. A usable image still needs your judgement.