ChatGPT Photo Editing Prompts: The Copy-Paste Library for GPT Image 2

Yes — ChatGPT edits photos with plain English, and the prompt that works is a two-part sentence, not a magic word. The engine behind it is GPT Image 2, Open​AI's image model that shipped April 21, 2026, and the workflow is upload a photo, highlight what you want changed, and describe the edit in a sentence. The formula that gets good results has three slots: change [what], keep [the rest] unchanged, [guardrail]. Below is the whole thing in copy-paste form — a fill-in-the-blank template, ten prompts for the edits people actually do, the step-by-step workflow, and the places where ChatGPT is the wrong tool. This page is specifically about editing images you already have (image-to-image); if you came for text-prompting patterns for G​PT-​5.6 — writing, coding, reasoning — that is a separate field guide, not this one.

The short answer: what you get, and what it costs

ChatGPT's photo editing runs on GPT Image 2, Open​AI's image model launched April 21, 2026, which now powers both generation and editing in the ChatGPT app (the chat layer above it is G​PT-​5.6, but the pixels come from GPT Image 2). You upload a photo into the chat, open the built-in editor, brush over the area you want changed, and type the edit. It works on the free tier — with limits that matter:

• Free ($0) — text chat became unlimited this month (Open​AI lifted the free text-chat cap the week of August 10, 2026), but image generation is still separately capped. Open​AI does not publish the number; independent trackers put it around 3–10 images per rolling three-hour window depending on load. You get roughly 500 MB of library storage and can upload up to three images at once for multi-image edits.

• Plus ($20/month) — adds Thinking mode (the model plans before it renders) and web search during generation, both of which visibly improve complex edits, plus higher caps and priority during peak load. This is the tier to buy if you edit photos daily.

• Pro ($100–200/month) — "unlimited and faster" image creation, subject to abuse guardrails, with access to the Sol Pro model tier for the chat layer.

• Resolution — up to 2,000 pixels on the long edge in the consumer app, across the full aspect-ratio set. 4K output exists but only through the API, and even there output above 2K is flagged experimental.

The fill-in-the-blank template: change + preserve + guardrail

Text-prompting guides talk about context, constraints, and examples — that is how you prompt a reasoning model like G​PT-​5.6 for text output. Editing a photo is different: the model already sees the image, so a generation-style prompt wastes your strongest signal. Edit prompts are built from one pair that carries nearly all the weight: what to change and what to keep unchanged, plus a guardrail that says how far the change may go.

Copy this and fill the brackets: Change [exactly what changes — one thing only]. Keep [everything else — lighting, pose, colors, background, identity] exactly unchanged. Amount: [subtle / moderate / strong]. Do not [the failure to block — over-smooth, add objects, change the face, add readable text].

Filled-in example: Change the background to a warm golden-hour beach with soft waves. Keep the subject's pose, clothing, skin tone, and expression exactly unchanged. Keep the original lighting direction on the person. Match the new background's light to the person's existing shadows. Do not add people, animals, or readable text.

The two "keep" clauses and the "do not" clause do more work than the "change" clause. GPT Image 2 is aggressive about filling ambiguity with its own guesses — every unsaid thing is a coin flip. The "keep" clauses are your contract with the model, and they are the difference between an edit and an accident.

The copy-paste library: ten prompts for the edits people actually do

Each of these is written to paste in one shot after you have uploaded a photo — and, for a local edit, brushed the region first. Swap the bracketed parts for your own.

1. Replace the background. Replace the background with [a misty pine forest at dawn]. Keep the person and their clothing, hair, and shadows exactly as they are. Match the new background's light to the person's existing shadows. No readable text.

2. Remove an object or a person. Remove the [person on the right]. Fill the space using the surrounding background texture and light. Keep everything else identical. No blur, no smudge, no new objects.

3. Natural portrait retouch. Even out skin tone and reduce blemishes. Preserve pores, skin texture, and facial structure — do not smooth or plasticize. Keep eye shape, expression, and clothing unchanged. Subtle.

4. Color grade for mood. Shift the color grade toward [warm autumn tones / cool teal / moody desaturated]. Keep brightness and contrast similar to the original. Apply the grade evenly to the whole frame. Do not push skin tones into an unnatural cast.

5. Fix a dark or underexposed photo. Brighten the image with realistic lighting. Recover detail in the shadows while protecting the highlights. Reduce noise in the darker areas. Preserve the original mood and color balance. Subtle.

6. Restore an old or damaged photo. Repair scratches, stains, folds, and dust. Reconstruct damaged areas from the surrounding detail. Keep the person's identity, expression, and the look of the original era. Add realistic, muted color where it is missing. Do not modernize or over-sharpen.

7. Sharpen and upscale a blurry photo. Reconstruct this low-resolution image at high fidelity. Recover edges and fine detail without introducing artifacts or noise. Keep the original lighting and color. Do not invent texture that is not in the source.

8. Film look / style transfer. Convert this photo to [1970s Kodak Portra film look]: slight grain, warm yellow-brown tones, a gentle vignette, softened shadows. Preserve the subject's features and realism. Keep the composition unchanged.

9. Clean product shot for a listing. Remove dust, smudges, and reflections from the [product]. Keep the product's color, logo, and proportions exact. Replace the background with seamless white. Add a soft natural shadow under the product. No props, no extra text, no watermark.

10. Remove a timestamp or watermark. Remove the [watermark in the corner]. Fill the space with the surrounding pattern and lighting so no trace remains. Keep everything else identical.

Two notes on the library. First, "keep … exactly as they are" is the phrase that prevents most drift — use it literally. Second, GPT Image 2 renders short text well (roughly under eight words per line) but longer strings break; for prompt 9 and any edit that must keep a logo or label legible, the "keep … exact" clause is what you rely on, and you should still verify the output zoomed in.

Step-by-step: the editing workflow that works

You do not need the brush for everything — a whole-image edit like a color grade or a film look can be typed straight into the chat after you upload. But for local edits (remove an object, swap a background, retouch a face), use the native editor:

• 1. Upload or generate. Click the paperclip / "+" in the chat box and attach the photo, or generate one in ChatGPT first and edit that.

• 2. Open the editor. Click the uploaded image. This opens the image editor with a Select (brush) button at the top right.

• 3. Brush the region. Use the size slider to match the brush to the area — small for a blemish, larger for a background. You can highlight several areas at once and undo/redo freely.

• 4. Type the edit. Paste the relevant template or library prompt into the box, keeping "keep everything else unchanged" for the parts you did not brush. Send.

• 5. Check, then iterate. Zoom in on the edited edge — the change often spills a little past the brush stroke. If something went too far, follow up with a small corrective prompt ("restore the original [feature], keep the rest") instead of repeating the whole edit.

• 6. Save / download. Save from the editor, or use the download arrow on the result in the chat.

The single most useful habit: change one thing per prompt, and say "keep everything else unchanged." When an edit fails, it fails because the model reinterpreted something you never mentioned. A one-slot change makes the failure legible, and the next prompt is a tiny correction instead of a full re-roll.

Where ChatGPT photo editing is the wrong tool

This is the section the listicles skip. ChatGPT's image editing is genuinely good — for one-off creative edits. There are five situations where it will cost you more than it saves.

1. Face and character consistency across many images. This is GPT Image 2's documented weak spot. Editing a single photo keeps the person recognizable, but editing or generating that same person across five shots drifts — the face shifts, the outfit changes, the logo mutates. For a consistent series (product hero shots, brand assets, a campaign with one model) this is the wrong tool. The 2026 "subject lock" improvements reduce the problem but do not remove it.

2. Exact brand colors and pixel fidelity. "Keep the logo exact" works until the logo is small or the color has to match a brand hex. The model is reconstructing, not measuring — it will happily approximate. If the deliverable has to match a swatch, verify every output against the original or use a pixel editor.

3. Long or dense text. Short text is a strength; anything past roughly eight words per line degrades, and fine print is out of scope. Do not use ChatGPT to redo a poster full of body copy.

4. Print and large-format resolution. The consumer app caps output at 2,000 pixels on the long edge — fine for web and social, thin for print. The 4K tier is API-only and flagged experimental above 2K.

5. Batch volume. Editing is one prompt per image, and every render carries a token bill. GPT Image 2 runs about $8 per million input tokens and $30 per million output tokens — a typical edit lands around $0.04–$0.25 per image depending on quality tier, up to roughly $0.41 for a 4K high-tier render (numbers from our own GPT Image 2 pricing comparison, which used Open​AI's published rates). A 500-SKU catalog at an average of $0.10 per image is a $50 render bill plus your review time. Cheaper than hand retouching, but slow to babysit, and only worth it when one prompt template applies across the whole batch.

There is also a content-policy boundary to know before you try. Open​AI blocks editing real identifiable people into fabricated or explicit scenarios, and celebrity or trademarked-character requests are refused. Legitimate edits of your own photos — removing a photobomber, restoring a family photo — are fine; the refusal you will actually hit is the edit that would put a real person in a false or harmful context. Rejected requests still count against your image cap, so phrase edits conservatively the first time.

When any of these five hit you, the fix is usually not a better prompt — it is a different pipeline. For one-off work, accept the limits. For consistent or high-volume work, move to the API where you control resolution, quality tier, and seed: GPT Image 2 is hosted on OrcaRouter at provider list price with 0% markup — same model and same Open​AI price, with automatic failover if the primary provider is down — and batch API pricing roughly halves the token cost. Every prompt in the library above works identically in the API, with a seed parameter added for reproducible renders.

Common pitfalls

• "Enhance this photo." The most common prompt and the least useful. It hands every decision to the model, and the result is over-sharpened skin and cranked saturation. Always name the specific edit.

• Stacking edits. "Change the background, make it warmer, remove the cup, and sharpen" fails in four places at once and you cannot tell which one broke. One change per prompt.

• Forgetting the keep clause. The single biggest cause of "it changed the wrong thing." Put "keep everything else unchanged" at the end of every local edit.

• Trusting the brush outline. The edit often spills past your stroke. Zoom in at the selection edge before saving.

• Hitting the free cap mid-job. Image generation stays capped on the free tier even though text chat is now unlimited, and rejected requests still count. Batch your edits into fewer prompts, or move to Plus for Thinking mode and priority.

• Skipping zoom. Faces, logos, and text all look right at thumbnail size and wrong at 100%. Zoom before you download.

Bottom line

ChatGPT photo editing is the closest thing most people have to a zero-cost retoucher: upload, brush, prompt, download — the free tier handles one-off edits, and the API handles real workloads. The skill is not prompt craft so much as constraint discipline: say what to change, say what to keep, and change one thing at a time. The template and the ten prompts above cover the edits people actually do. Keep the "keep everything else unchanged" habit, zoom before you save, and move to the API the moment you need consistency or volume — where GPT Image 2 routes on OrcaRouter at list price with failover, and the same prompt library works with a seed for reproducibility.