Fastest fix: click your image, pick the Select tool (the paintbrush in the image toolbar), brush over only the area you want changed, then describe the change. That scopes the edit to one region and stops ChatGPT from regenerating the whole picture into a near-identical copy. If you can’t or don’t want to use Select, the next-best move is a three-part keep / change / avoid prompt (Step 2). Those two together fix the large majority of “nothing changed” cases.
Here’s why this happens. ChatGPT’s in-app image editing (the GPT Image series — gpt-image-1.5 is the production default as of June 2026, with gpt-image-2 rolling out since April 21 2026) is image-to-image regeneration, not true pixel-level editing of your original file. You upload an image, the model reads it via vision, then generates a new image “that looks like the original but with your edit applied.” Give it a whole-image text prompt with a vague instruction and it resolves the ambiguity the safest way — preserve everything — so you get back a picture nearly identical to the one you started with. The Select tool (added through the 2025-26 “ChatGPT Images” updates) sidesteps this by telling the model exactly which pixels are in play.
DALL·E 2 and DALL·E 3 were retired on May 12 2026; if an older guide tells you to “switch to DALL·E 3,” that path no longer exists. All in-app image work now runs through the GPT Image models.
Which bucket are you in?
Ordered by hit rate, highest first. Match your symptom to a cause, then jump to the matching step below.
| # | Cause | Tell-tale sign | Fix |
|---|---|---|---|
| 1 | Edited the whole image with a text prompt | You typed the change into chat without using Select; result is a near-identical full redraw | Step 1 (Select tool) |
| 2 | Vague verbs (“edit”, “tweak”, “fix it up”) | No keep / change / avoid structure in the prompt | Step 2 |
| 3 | Didn’t say what to preserve | Subject pose / color drifted, or nothing changed at all | Step 2 |
| 4 | Source image too low-res | Original long edge < 800px | Step 3 |
| 5 | Edit targets text / logo / numbers | The thing to change is rendered text, digits, or a logo | Step 5 |
| 6 | Safety layer quietly downgraded the edit | Request is “make X look more Y”, a celebrity, or a sensitive theme | Step 6 |
| 7 | Multi-turn drift | First edits were fine, detail decays from the 3rd round on | Step 7 |
1. You edited the whole image instead of a region
The single biggest cause now that the Select tool exists. Type “change the background to blue” into chat and the model regenerates the entire picture, then defaults to “preserve everything” — so you get a near-identical copy. Brush only the background with Select and the model has no reason to touch the subject.
How to spot it: you described the change in the chat box, not by selecting an area first.
2. Prompt uses vague verbs (“edit,” “tweak,” “fix it up”)
“Make this image look nicer” reads to the model as “I don’t really know what you want, safest is to leave it alone” — so it returns a near-identical image.
How to spot it: Your prompt has no “keep X / change Y / don’t touch Z” structure = too vague.
3. Didn’t specify what to preserve
“Replace the background with blue” — the model doesn’t know whether to preserve the foreground. It might redo the subject too, or freeze both to be safe.
How to spot it: Returned image has subtle changes to subject pose / color / detail (or no changes at all) = preserve signal missing.
4. Source image too low-res; model can’t isolate regions
Vision struggles with small images. If it can’t tell “what is background vs subject,” it regenerates the whole image holistically — usually nowhere near your target.
How to spot it: Original long edge < 800px = likely resolution issue. Re-upload at 1500px+ and retry.
5. Edit involves specific text / logo / numbers
Precise control of rendered text inside images is still a known weakness of this model generation as of June 2026 — non-Latin scripts (Chinese, Arabic, Hebrew) and small/dense text are especially fragile. “Change the price from $99 to $79” often returns garbled characters or leaves the text unchanged.
How to spot it: The thing to change is text / digits / a logo = current model capability boundary. Use the Select tool for a clean region, or a real editor (Photoshop / Figma).
6. Safety filter silently weakened the edit
Edits that touch facial features, unusual poses, or sensitive themes can quietly get routed into a “conservative edit” path — returning a barely-changed image.
How to spot it: Edit request involves “make X look more Y” (younger / taller / thinner) / celebrities / political symbols = likely safety-layer downgrade.
7. Multi-turn drift in the same chat
By the third edit on “the same” image, the model is actually editing v2 (its previous output), not the original. Small losses accumulate per round.
How to spot it: First two edits looked right; third onwards detail starts drifting = multi-turn drift. Restart from the original each round.
Before you start
- Confirm whether this happens in a plain chat or a Custom GPT — image limits differ across plans.
- Back up the chat and the original image before retesting so history doesn’t pollute the next diagnostic.
- Check your quota. As of June 2026: ChatGPT Free gets roughly 2-3 image generations per rolling day; Plus ($20/mo) gets about 50 images per 3-hour window; Pro ($200/mo) is effectively unlimited (subject to abuse guardrails). On Free, an over-quota request can fail or return a stale image rather than a clear error — if you’ve generated a few images already today and edits stopped landing, you may simply be rate-limited.
Info to collect
- Source image resolution (W × H), file size, origin (own photo, web image, AI-generated).
- Full prompt text + returned image screenshot (ideally side-by-side with original).
- Concrete description of expected vs actual difference (“sky should be blue, came back white”).
- Current model + whether in plain chat / Project / Custom GPT.
Shortest fix path
Ordered by ROI. The first two solve the large majority of cases.
Step 1: Use the Select tool to scope the edit to one region
This is the highest-leverage fix and the one most older guides miss. Instead of describing a change in chat (which redraws the whole image), tell ChatGPT exactly which pixels to touch:
- Click the image in the chat to open it full-size.
- In the image toolbar, choose Select (the paintbrush / region-selection icon; sometimes called the in-painting or “Edit” tool).
- Drag the brush-size slider to a brush that comfortably covers your target area, then paint over only the region you want changed (e.g. just the sky). Use Undo / Redo to clean up the selection.
- With the area highlighted, type the change in the prompt box — for example “make this a clear bright-blue sky with scattered white clouds.”
- The model regenerates only the masked region and leaves the rest untouched.
This works on both AI-generated and uploaded photos. It is the most reliable way to change a background, remove an object, or swap one element without the subject silently drifting. It’s available on web and the mobile apps for paid plans; if you don’t see it, update the app and confirm you’re on a plan with image editing.
If your edit is genuinely global (restyle the whole image, change overall lighting), skip Select and go straight to the structured prompt below.
Step 2: Use a three-part keep / change / avoid prompt
For whole-image edits, convert vague requests into structured instructions:
Edit this image:
KEEP:
- The person's pose, facial features, and clothing exactly the same
- All details in the foreground unchanged
CHANGE:
- Replace the background sky from overcast grey to bright blue with
scattered white clouds
- Add soft warm sunlight from the upper right
AVOID:
- Do not alter the person at all
- Do not add or remove any objects
- Do not change the lighting on the foregroundThe three-part structure makes the model handle preservation and modification as separate concerns, which alone fixes many “nothing changed” results.
Step 3: Push resolution to >= 1024px
Low resolution → vision can’t see clearly → region detection fails → holistic regenerate. Re-upload:
- Phone originals (usually 3000px+), not thumbnails.
- Upscale screenshots first (macOS Preview → Tools → Adjust Size → 1500px on long edge).
- Keep AI-generated images at the original 1024px+ without recompression. GPT Image outputs are png/webp/jpg under 50MB, so size is rarely the blocker — sharpness is.
Step 4: Two-step — describe first, then edit
When detection is weak, split the task:
Turn 1: Describe everything in this image in detail — subject, pose,
clothing, background, lighting, composition.
Turn 2 (after it answers): Good. Now edit only the background — change
it from <its description> to <your target>. Keep everything else
exactly as you described.Its own description is more precise than yours — use it as the “preservation contract.”
Step 5: For text / logo edits, switch tools
Accept the current model boundary. As of June 2026, GPT Image handles short Latin text reasonably but is still unreliable on small/dense text, exact digits, and non-Latin scripts (Chinese, Arabic, Hebrew). Precise text / numbers / logo work belongs in:
- Simple text: Canva / Figma / Photoshop text tool.
- Complex posters: Photoshop generative fill (more reliable).
- QR / barcodes: re-generate locally with a real tool — never let the model “redraw” a scannable code.
Don’t waste three rounds arguing with ChatGPT about this. If you only need to swap text around an existing image, the Select tool (Step 1) plus a layout app beats fighting the prompt.
Step 6: Rewrite sensitive edits as neutral
For “make X look Y” or feature changes:
Bad: Make her look 10 years younger.
Good: Put her in a different outfit — replace business suit with
casual sweater and jeans. Keep face, hair, body proportions
unchanged.Reframe “edit the person” as “edit the styling / environment” — much less likely to trigger safety downgrades.
Step 7: Restart from the original every round
Don’t iterate 5 rounds on one image. Before every major change:
- Download the original (or the best version so far).
- Open a new chat and re-upload that as the source.
- Accumulate prior successful changes into the new prompt.
Avoids multi-turn drift.
How to confirm the fix
- Open a fresh chat, upload the same original, re-run the Select-tool edit (or the rewritten three-part prompt) — output changes as expected = truly fixed.
- Verify each “preserve” item one by one (face, pose, clothing, foreground objects) — model genuinely left them alone = edit precision is real.
- Overlay the result on the original at 100% in any viewer; the masked region should differ and everything outside it should be pixel-stable. If untouched areas also shifted, you edited globally — switch to Select.
- Have a colleague run the same prompt in their account — output direction matches = prompt is general, not your lucky roll.
If still broken
- Cut to the simplest test: 500×500 solid background color change with Select — confirms base capability still works.
- Swap image source: own photo → AI-generated → web image — rules out safety triggers from a specific source.
- Check status.openai.com — image generation degradations and partial outages show up there before they’re obvious in-app.
- Update the app / hard-refresh the web client; the Select tool and current GPT Image model ship through client updates, and an old build can leave you on the whole-image-only path.
- Fallback path: route the edit to a dedicated tool (Canva / Photoshop generative fill / Krea), or use an off-OpenAI image-to-image model like Flux Kontext / FLUX Edit.
Prevention
- Reach for the Select tool first for any localized change (background, one object, a region); save text prompts for genuinely global edits.
- For whole-image edits, use the three-part keep / change / avoid structure — never natural-language paragraphs.
- Source resolution >= 1024px, the larger the better; upscale low-res first.
- Don’t expect ChatGPT to be precise about text / logos / numbers — use Photoshop / Figma.
- For multi-round edits, restart from the original each round to avoid v1 → v3 quality loss.
- For high-stakes commercial images, generate 3 variants (“give me 3 options to choose from”) rather than betting on one output.
FAQ
Why does ChatGPT return the same image when I ask for an edit?
Because a whole-image text prompt is regenerated from scratch, and an ambiguous instruction is resolved by preserving everything. Use the Select tool to mask just the area you want changed, or write a keep / change / avoid prompt so the model knows what to alter.
Where is the Select / paintbrush tool in ChatGPT? Click the image to open it full-size; the Select (region/paintbrush) icon is in the image toolbar. Brush over the area, adjust the brush-size slider, then type the change. If you don’t see it, update the app — it ships with the current “ChatGPT Images” client and requires a plan with image editing.
Does ChatGPT edit my actual file, or make a new image? It generates a new image. The in-app editor is image-to-image regeneration (powered by the GPT Image models), not pixel-level editing of your uploaded file. The Select tool limits how much of that new image diverges from the original.
Can I edit images on the free plan? Yes, but Free has a tight quota — roughly 2-3 generations per day as of June 2026 — and over-quota requests can fail quietly. If edits stop landing after a few tries, you may be rate-limited; wait for the window to reset or use Plus/Pro.
Why won’t it fix the text in my image? Rendering exact text — especially small/dense text, precise digits, or non-Latin scripts like Chinese, Arabic, and Hebrew — is still a weak spot of this model generation. Do text in Canva / Figma / Photoshop and use ChatGPT only for the surrounding imagery.
Is DALL·E 3 still an option for editing?
No. DALL·E 2 and DALL·E 3 were retired on May 12 2026. All in-app image generation and editing now runs through the GPT Image series (gpt-image-1.5, with gpt-image-2 rolling out).