Three ChatGPT photo prompts that fix most bad results

Short answer: The three instructions that most improve ChatGPT photo results are (1) say what must stay the same, not only what should change, (2) put any text you want printed in quotation marks, and (3) describe the shot, lens, light, distance, instead of the subject alone. All three matter more since ChatGPT Images 2.5, released on 8 September 2026.

What changed in September 2026

OpenAI introduced ChatGPT Images 2.5 with sharper detail, more natural lighting and texture, better preservation of the people and products in your reference photos, more reliable multi-turn editing, and up to 50% faster generation than the previous model. It also added Sketch (draw a rough layout as a reference), templates for common formats, and comments you can place directly on an image to fix one area. For developers there are two API models: a speed-first one and a higher-precision one for detailed editing.

The practical shift: the model is now good at controlled transformation. So your prompt should describe the destination and what must survive the trip.

Code 1, The preservation line

Most disappointing edits happen because the model changed something you never mentioned. Add an explicit preservation clause.

Edit the uploaded photo. Change only the background to a plain warm-grey studio wall.
Keep the person's face, hair, skin texture, clothing and body position exactly as in
the original. Do not restyle, smooth or re-light the subject.

Use it whenever a face, a product or a logo appears.

Code 2, Quoted text

Anything you want printed on the image goes inside quotation marks, spelled exactly. Unquoted words get treated as a theme, not a requirement.

A minimal product card, warm stone background. Printed text, exactly as written:
headline "Autumn Edit", small line beneath "New arrivals, Friday".
No other text anywhere in the image.

Add "no other text anywhere", it prevents invented captions.

Code 3, The camera line

Describe it as a photographer briefing a shoot, not as a shopper describing a thing.

Full-length shot, 50mm lens, eye level, subject centred with space above the head.
Single soft window light from the left, soft shadow to the right, no fill.
Muted natural colour, fine film grain, no glossy retouching.

Distance, lens, light direction and grain do more for realism than any adjective like "professional" or "high quality".

Put the three together

Edit the uploaded photo of the product.
KEEP: the product shape, label text, colour and material exactly as photographed.
CHANGE: place it on a warm stone surface with a soft shadow.
CAMERA: 85mm, slight top-down angle, single soft light from upper left.
TEXT: none.

Then refine in the same chat, one change per turn. Multi-turn consistency is the thing Images 2.5 improved most, so small corrections now beat rewriting the whole prompt.

Common mistakes

• Asking for five changes in one message. Do one, confirm, then the next.
• Writing "make it look professional". It means nothing to the model, name the light.
• Forgetting the negative. "No watermark, no extra limbs, no invented text" is short and it works.
• Uploading a low-resolution reference and expecting detail the source never had.

Frequently asked

Which model is behind this? ChatGPT Images 2.5, available to ChatGPT users on desktop, mobile and web since 8 September 2026.

Can it keep the same person across several images? It is much better at it than earlier versions, but still not guaranteed. Reuse the same reference photo and the same preservation line every time.

Can I use the results commercially? Check your plan's current terms and your client's policy on AI imagery, and never generate a recognisable person without permission.

Sources: OpenAI's ChatGPT Images 2.5 announcement (8 September 2026) and OpenAI's image generation documentation.