ChatGPT’s image editing (DALL·E 3 / GPT-Image generation) is actually image-to-image regeneration, not true pixel-level editing. You upload an image, the model glances at it via vision, then generates a new image “that looks like the original but with your edit applied.” Any vague instruction gets resolved by the safest path — preserve everything — and you get back an image nearly identical to the original. The fix isn’t waiting for a smarter model; it’s writing prompts the model can execute unambiguously.
Common causes
Ordered by hit rate, highest first.
1. Prompt uses vague verbs (“edit,” “tweak,” “fix it up”)
The most common failure. “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.
2. 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.
3. Source image too low-res; model can’t isolate regions
Vision struggles with images under ~512px. 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.
4. Edit involves specific text / logo / numbers
Precise control of rendered text inside images is a known weakness of this generation of models. “Change the price from $99 to $79” usually returns gibberish or unchanged.
How to spot it: The thing to change is text / digits / a logo = current model capability boundary. Use Photoshop instead.
5. 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.
6. 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 generation call limits differ across Free / Plus / Pro.
- Back up the chat and the original image before retesting so history doesn’t pollute the next diagnostic.
- Confirm your plan: Free users have tight image-edit quotas — over-quota requests can fail silently.
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 ~70% of cases.
Step 1: Use a three-part keep / change / avoid prompt
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 foregroundMassive quality jump. Three-part structure makes the model handle preservation and modification as separate concerns.
Step 2: 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.
Step 3: 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 4: For text / logo edits, switch tools
Accept the current model boundary — 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.
Don’t waste three rounds arguing with ChatGPT about this.
Step 5: 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 6: 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 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.
- 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 — confirms base capability still works.
- Swap image source: own photo → AI-generated → web image — rules out safety triggers from a specific source.
- Try different image models: 4o image vs GPT-Image vs DALL·E 3 (depending on what’s available to you).
- Fallback path: route the edit to a dedicated tool (Canva / Photoshop / Krea), or use an off-OpenAI image-to-image model like Flux Kontext / FLUX Edit.
Prevention
- Always use the three-part keep / change / avoid structure for image edits — never natural-language paragraphs.
- Source resolution ≥ 1024px, the larger the better (up to 4096px); 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.