AI Photo Editing Prompts: Why Most Fail — and the Five Parts of One That Works

Everyone has run this experiment. You upload a photo of yourself, type "enhance this photo, make it professional, ultra realistic, 8k", and get back something that is undeniably sharper — and undeniably not you. The jaw is narrower. The skin is candle wax. The shirt has been redesigned, and the room behind you has been replaced with a vaguer, more confident room.

The conclusion most people draw is that the model is bad at faces, or that there is a secret vocabulary they were never given. Both conclusions are wrong, and the real explanation is more useful: the prompt was not too short — it was missing entire jobs.

The mistake: treating an edit like a creation

A prompt that creates an image only has to describe what you want to see. A prompt that edits a photo has a second, larger responsibility: describing what must not change. That second half is the one everybody skips, and it is where every ruined edit comes from.

Image models do not repair photographs; they re-imagine them. Left without instructions, they drift toward what their training says photos usually look like — and the drift always flows the same direction: faces slim, teeth whiten, skin smooths, backgrounds tidy themselves, clothes get restyled into something a stock-photo model would wear. None of this is the model misbehaving. A five-word prompt simply granted it permission to change everything, and it used the permission.

So the skill of AI photo editing is mostly the skill of forbidding things precisely. Which sounds restrictive, but is exactly how you would brief a human retoucher: "clean up the blemishes, but I want to still look like me" is a lock list. The model needs the same brief, in more detail, because it has less judgement.

The five parts of a working edit prompt

Every reliable edit prompt we have written — and we have written a lot of them, tested across the current models — ends up with the same five blocks, in roughly the same order.

1. An edit verb, first. Open with retouch, re-light, rebuild, restore, re-photograph — never "make a picture of" or "generate". An edit verb tells the model the uploaded photo is the subject of the work, not loose inspiration for it. This one word changes how literally the model treats your reference.

2. The lock list. Everything that must survive the edit, stated explicitly: identity, facial geometry, apparent age, expression, gaze, pose, crop, camera angle, clothing, background, lighting — whichever of those the task does not deliberately change. This is the highest-value sentence in any edit prompt and the one almost every prompt you will find online is missing. If the result must still be the same human being, say exactly that: "the output must be recognisable as the same individual".

3. The build list. The change itself, described in physical terms rather than wishes. "Make it look professional" is a wish — the model fills the vagueness with stock-photo clichés. "Replace the background with plain studio grey and dress the subject in a navy crew-neck" is an instruction. The rule of thumb: describe materials, geometry and surfaces, not qualities.

4. Camera and light. Photographers' language moves these models more than any adjective, because photographs in their training data came captioned with it. "85mm at f/2.8, one large soft key just above eye level" reliably beats "professional studio photo". You do not need to know photography — you need about ten of these phrases, and you can steal them from any lighting tutorial.

5. The negative list. Name the specific failure modes for the task, out loud, at the end: waxy skin, warped hands, extra fingers, halo edges around hair, invented text, HDR crunch. Naming a failure measurably suppresses it. Generic negatives ("no artifacts") do little; specific ones do a lot.

A worked example: from five words to a real prompt

Take the classic job — turning a decent phone snap into a headshot for a profile. The five-word version, "make this a professional headshot", produces the stranger from the opening paragraph. Built block by block instead:

Re-photograph the supplied portrait as a corporate headshot. Keep the same person exactly — identity, facial geometry, apparent age, expression, skin texture and hair must be unmistakably the same individual; do not slim the face, whiten the teeth, enlarge the eyes or add makeup. Replace the background with plain soft-grey studio paper, and dress the subject in a plain navy crew-neck. Light it like an 85mm portrait at f/2.8: one large soft key just above eye level, gentle fill, a natural catchlight. Skin keeps its pores and fine lines — matte, no gloss, no beauty filter. Avoid: waxy skin, plastic smoothing, warped ears, halo edges around hair, a different person.

Count the jobs: one edit verb, four sentences of locks and builds, one line of camera language, one negative list. About a hundred words — and roughly half of them are there to stop things from happening. That ratio is not padding. On current models, it is the difference between an edit and a replacement.

The five failures behind almost every bad result

When a well-built prompt still misfires, it is nearly always one of these — and each has a known fix. Add the fix to the prompt and regenerate from a fresh upload; arguing with the model in follow-up messages compounds the drift, because each round re-imagines the previous round's output.

• The face is no longer them. Your source face is too small — drift is worst when the face occupies few pixels. Crop tight on the face, enhance that first, then use the result as the reference. And add the "same individual" clause from the lock list.
• Skin looks like wax. Delete every word like smooth, flawless, perfect — they are beauty-filter triggers. Ask for the opposite: "retain visible pores, skin texture and fine lines; matte finish".
• The effect is far too strong. State an intensity: "apply at roughly 40 percent strength". Nearly every current model respects a stated percentage, which surprises everyone the first time.
• The background got reinvented. You did not lock it. "Match the layout, spacing and geometry of the reference exactly; do not reinterpret or reposition anything."
• Text turned to gibberish. Lettering is the hardest thing in the field. Type the exact words into the prompt — "the sign reads exactly: OPEN DAILY" — and use the strongest model available to you; most simply cannot hold text at all.

Where to run prompts for free

Allowances change frequently, but as of writing the ranking is clear. Google AI Studio (aistudio.google.com) is the best free option by a distance: a proper daily allowance of capable image models, higher output resolution, and no visible watermark. The Gemini app is more convenient but its free tier allows only a couple of top-model generations a day, watermarked. ChatGPT's free tier includes a small daily allowance of edits and is the more forgiving model for painterly, stylised transformations — though it sometimes declines edits of real people's photos, which is a policy filter, not a prompting failure.

Two honesty notes. There is no free API route — the capable image models are paid-only for developers, so the free allowances live in the web interfaces, not in code. And even an unwatermarked output carries invisible provenance metadata identifying it as AI-generated. It does not change how the image looks or your right to use it, but it is detectable, which is worth knowing before you use one somewhere that matters.

When a prompt is the wrong tool entirely

One job deserves a warning label: restoring damaged family photos at volume. A generative model re-imagines a face rather than repairing it, so likenesses drift — which is precisely the wrong failure on the only photo you have of your grandmother. Add daily caps and the fact that every upload lands on someone's server, and a shoebox of three hundred family photos is a bad match for a chat window. For one or two photos, a careful restoration prompt is genuinely good now; for an album, use a tool that runs the restoration locally on your own computer instead — we wrote a full guide in how to restore old photos at home, and our own PhotoFix exists for exactly that shoebox case: no uploads, no per-photo credits.

Fifteen prompts already built this way

If you would rather start from working prompts than write your own, we published The AI Photo Lab — fifteen edit prompts built on exactly the anatomy above: detail-recovering upscale, invisible retouch, old-photo rescue, background swaps, headshots, product shots, and the scene rebuilds people share. Each one has a copy button, the fill-in parts marked in brackets, and a note on which model it was written for. It is free on the page or as a PDF, with no signup — and if AI beyond prompts is new territory, the free AI Mastery course starts from zero, in English, Romanian and Italian.

AI models and their free tiers change without notice, so specifics here will age — the five-block anatomy is the part that transfers. Images you generate from your own photos are yours; check the current terms of whichever service you use for anything commercial.