AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
Most people ask an image tool for "a professional headshot of me" and attach one holiday photo. What comes back is a polished stranger with your hair colour: the jaw is tidier, the eyes are a different shape, and the skin looks airbrushed. It feels fine for ten seconds, then a colleague says "that doesn't look like you" and the picture is useless on a profile you actually want people to recognise.
You can get something much closer with ChatGPT and photos you already have, no extra app, connector or plugin. But there is one honest limit, covered below, and it changes how you should use the results.
The rule: give ChatGPT several honest reference photos, tell it exactly what to keep and what to change, and judge every result against your real face, not against how polished it looks.
What you need
This prompt tells the model what must stay fixed (your face) before it describes what may change (clothes, backdrop, light). Fill in the square brackets.
You are an experienced portrait photographer and retoucher. I have attached [NUMBER] photos of myself. Treat them as the single source of truth for my face.
Goal: produce a professional headshot that people who know me would recognise instantly. A good result looks like a well-lit photograph of me on a good day, not a generic stock model.
Keep exactly as in my photos: face shape, eye colour and shape, nose, mouth, [DISTINCTIVE FEATURES, e.g. glasses, freckles, beard, grey at the temples], skin texture, hairline and hair colour [HAIR NOTES].
Create these [NUMBER OF IMAGES] variations, one image each:
1. [OUTFIT 1, e.g. smart-casual shirt], straight-on to camera.
2. [OUTFIT 1], three-quarter angle.
3. [OUTFIT 2, e.g. dark suit jacket and shirt], straight-on.
4. [OUTFIT 2], three-quarter angle.
Style for all: soft, even studio lighting; plain [BACKGROUND, e.g. light grey] backdrop; head and shoulders; relaxed, friendly expression with a natural smile, not a fixed stare; [ASPECT RATIO, e.g. 4:5 portrait].
Constraints: do not retouch skin, slim the face, change my age, or alter any feature listed above. Do not add jewellery, text or logos. If my photos are not enough to keep a feature accurate, ask me which extra photo you need before generating.
Before generating, silently check that each planned image keeps every listed feature. Then generate.
If you want fewer images, say so. Generating one at a time lets you correct as you go.
Open each result beside your original photos and run this short checklist:
| Check | Pass if |
|---|---|
| Eyes | Same colour, spacing and shape |
| Hairline | Same height and shape, no thickened or invented hair |
| Jaw and cheeks | Not slimmed or sharpened |
| Skin | Visible texture, no plastic finish |
| Distinguishing marks | Glasses, freckles, scars or beard still present and in the right place |
| Teeth | Not whitened or reshaped beyond your own |
When a result is close but off, reply in the same chat naming one fault and what to preserve, for example: "Keep everything, but my hairline is too low; match the first photo." You can also open the image and use the editor: select Select, highlight the area, and describe the change, or describe the edit without the selection tool. OpenAI says highlights are not always precise and edits may extend beyond the area you selected, so recheck the whole face after any local edit.
Open the image and use Save to download it. Images you create are kept under Images; deleting an image from there means deleting the conversation it came from.
The trap is likeness drift. Each regeneration is a fresh attempt, not a retouch of your face, so small differences creep in, and a result can look flattering while quietly becoming someone slightly different. The honest limit follows from that: ChatGPT Images produces a likeness, not a guaranteed match, and I found no OpenAI statement promising identity accuracy. Treat the output as a good draft that you verify, never as proof of what you look like.
Two further facts worth knowing. OpenAI states that supported images generated with ChatGPT carry C2PA metadata and SynthID watermarks, so the images are labelled as AI-generated in ways other tools can detect. And some platforms or employers have rules on AI-altered profile photos, so check before you upload.