The Loop That Beats Luck
Most people generate 50 images hoping one lands. You're going to do the opposite: generate fewer, look harder, and change one thing at a time. Iteration is not "spam the button until magic happens." It's a tight loop — generate, critique, refine — where each pass is a deliberate experiment.
The mindset shift: treat your first image as a draft, not a verdict on the prompt. Drafts exist to be marked up. If you would not ship a first-draft essay, do not ship a first-draft render.
A good loop looks like this:
- Generate a small batch (4 images, not 40).
- Pick the one closest to your goal — not the prettiest, the closest.
- Write down what is wrong with it in plain words.
- Change exactly one thing in the prompt or settings.
- Generate again.
Five to ten passes of this gets you further than fifty random rolls.
Critique Before You Re-Prompt
The single biggest skill in iteration is honest critique. Before you touch the prompt, force yourself to name the gap between what you got and what you want. Vague reactions ("meh", "not quite right") produce vague edits. Specific reactions produce surgical ones.
Run every draft through a quick checklist:
- Subject: Is the thing you wanted actually the focal point? Right pose, right expression, right count?
- Composition: Where does your eye land first? Is the framing what you asked for?
- Lighting: Direction, hardness, color temperature — match your intent?
- Style: Does it look like the reference world you described, or did it drift toward generic "AI render"?
- Detail: Are hands, text, edges, and small objects clean, or quietly broken?
Write the critique as a single sentence: "Subject is right, but the camera is too high, lighting is too soft, and the background is too busy." Now you have three concrete things to fix — and a reason to fix them in a particular order.
Prompt Surgery: Change One Thing
The temptation is to rewrite the whole prompt every iteration. Resist it. When you change five things at once and the image improves, you do not know which change helped. You are gambling, not learning.
Treat the prompt like code. Make the smallest diff that addresses your top critique. If the lighting is wrong, edit only the lighting clause. If the composition is off, edit only the framing words. Keep a tiny log — even just a scratch note — of what you changed and what happened.
Here is a typical surgery sequence. Starting prompt:
A young chef plating a dessert in a restaurant kitchen,
warm light, candid, 35mm photo
Draft 1 looks too staged. Fix the candor:
A young chef plating a dessert in a restaurant kitchen,
warm light, caught mid-motion, slight motion blur on hands,
35mm photo, documentary style
Draft 2 nails the energy but the background steals attention. Fix the depth:
A young chef plating a dessert in a restaurant kitchen,
warm light, caught mid-motion, slight motion blur on hands,
35mm photo, documentary style, shallow depth of field,
background softly out of focus
Two clean edits, two clear wins. That is prompt surgery — not prompt rewriting.
Seeds, Variations, and When to Use Each
Modern tools give you two complementary levers: seeds (the random starting point) and variations (close cousins of an existing image). Knowing when to use each is the difference between productive iteration and chasing your tail.
Lock the seed when you have an image you mostly like and want to test prompt changes in isolation. Same seed + tweaked prompt = you can actually see what your edit did. Without locking, you are changing two variables — your prompt and the random noise — and any difference is ambiguous.
Use variations when the composition is right and you want subtle alternates — slightly different expressions, lighting, or details — without exploring a whole new image.
Reroll with a fresh seed when the image is fundamentally wrong and no amount of prompt nudging is rescuing it. Sometimes the seed itself is a dead end. After two or three failed surgeries on the same seed, abandon it.
A practical default: generate four images, pick the strongest, lock its seed, do two to three rounds of prompt surgery, then either ship or pivot to a fresh seed. If you are not closer after three locked-seed iterations, the problem is in your prompt's architecture (chapter 4) or your style references (chapter 5), not the seed.
Knowing When to Stop
Iteration has diminishing returns. The first three passes usually move the image dramatically. Passes four through seven polish it. Anything past ten and you are usually either lost or perfectionist — and perfectionism in generation often means you are trying to fix in-prompt what you should fix in post.
Three honest stop signals:
- Good enough for the use case. A thumbnail viewed at 200 pixels does not need pore-level detail. Match your effort to the surface it ships on.
- You are oscillating. Each iteration trades one problem for another. That is a sign the prompt has hit a ceiling — switch to inpainting or compositing instead.
- You have lost the thread. If the original goal feels fuzzy now, stop generating and re-read your own brief. Then start a fresh prompt informed by what you learned.
The next stage, where prompts hit their limit, is editing and inpainting — fixing the one bad hand, swapping the background, cleaning up text. That is a different toolkit and worth its own chapter.
A Concrete Drill
Do this once and your loop will sharpen permanently. Pick a single subject — say, "a vintage red bicycle leaning against a stone wall at golden hour." Give yourself a hard ceiling of ten generations. After each batch of four, write one sentence of critique and make one focused edit. At the end, look at your ten outputs side by side. You will see your own taste develop in real time, and you will spot the two or three edits that moved the needle most. Those become reusable moves in every project that follows.
If you want a structured warm-up before you do this, the AI Image Generation for Beginners course walks through the basic generate-critique-refine loop with shorter exercises, and AI Tools Comparison 2026 is useful when you suspect the tool — not your prompt — is the bottleneck.