Most people start with a blank prompt box and hope for the best. The faster route is sitting right there in your camera roll or your saved folder: an image you already like. The trouble is that uploading it and saying "make one like this" rarely gets you far. An AI tool does not see what you see. It reads pixels and turns them into statistics, and the parts you care about — the mood, the angle, the way the light falls — are not labelled anywhere.
A better habit is to use the reference as something to read rather than something to copy. You look at it deliberately, write down what actually makes it work, and hand those notes to the tool on the AI image generator. It takes maybe ten minutes once you get the hang of it, and it changes the odds a lot.
Break the reference into four layers
Every image can be described in four fairly separate layers, and it helps to keep them apart in your notes.
Subject is what is in the frame and what it is doing. "A woman in a wet coat, walking away, hands in pockets" is subject. It should take one line.
Composition is where things sit. Eye level or low angle? Tight crop or wide? Subject left of centre with empty space on the right? Is there a foreground element softening the bottom of the frame? Composition is the layer people skip, and it is usually the layer that makes a reference feel the way it feels.
Light covers direction, hardness and colour. Overcast light from above is flat and cool. Late sun from the left edge gives long shadows and warm skin. A single window with the rest of the room dark gives a bright cheek and a deep socket. Write the light down before you write anything about style.
Palette and texture are the finicky parts. Is the image mostly desaturated with one warm accent? Does the surface look grainy and analogue, or clean and digital? Are the shadows greenish or blueish? These are the details that separate a result that feels close from one that feels generic.
Write the prompt from your notes, not from the image
Once you have four short lines, build the prompt in a fixed order: subject, composition and camera, light, palette and texture. Keeping the order stable means that when a result is wrong, you know which line to change. If the subject is right but the mood is off, you edit the light line and leave the rest alone.
Try something like: "A woman in a wet wool coat walking away down a narrow street, eye level, slightly wide, subject left of centre; flat overcast light, cool shadows; muted grey-green palette, fine grain, 35mm look." It is not poetic. It is specific, and specific is what the model responds to.
Do not describe the thing you do not want
People often load a prompt with exclusions. "No text, no watermark, no extra fingers." The trouble is that naming something brings it into the frame, and a long list of negatives crowds out the description you actually need. Say what should be present, in as much concrete detail as you can, and use a short list of exclusions at most.
If a specific problem keeps appearing, that is a sign your positive description is thin in that area. Hands coming out wrong usually means you never described the hands.
Use the reference to check, not to grade
Once you have a result, put it beside the original and compare layer by layer. Ask four questions in order: is the subject doing the same thing, is the framing similar, does the light come from the same place, does the colour feel related. Whichever question gets a no is the line you edit. This sounds slow, but it removes the guesswork that makes people regenerate twenty times in a row and give up.
It also keeps you honest. A lot of "the AI can't do this" complaints are really "I described something different from what I pictured." Comparing layers catches that mismatch quickly.
When a reference is not the right starting point
Some references are actually poor teachers. A heavily filtered social photo has crushed shadows and a colour cast you will spend ages fighting. A screenshot of a compressed JPEG has blocking that the model may faithfully reproduce as texture. A famous photograph will pull the output towards that specific image rather than towards the qualities you liked in it.
Pick references that are technically clean and not too iconic. If the only good example you have is famous, describe it in words and skip the upload — you will get more variety and less accidental imitation.
Build a small library of notes
After a few sessions you will have a stack of four-line descriptions. Keep them. They become a personal shorthand: you can paste a light line you know works, swap the subject, and get something new in one pass. That is a much better long-term setup than starting cold every time.
If you want to try this out, open the best AI image generator, pick one image you genuinely like, and write the four lines before you type anything else. The first attempt will probably still miss. The third one usually does not.