The short answer: yes — in our test, every time. We uploaded one portrait to the Gemini app and made 26 photo edits: weddings, a roadside chai stall, the Hunza river, and our own five most-copied prompts. First try only. No re-rolls, no picking the best of several. Then we scored every photo trait by trait against the original face. All 26 kept the face.
The best wording was also one of the simplest:
Keep the exact same face as the uploaded photo: same eye shape and spacing, nose width, lip shape, jaw, cheekbones, skin tone, and age. Do not beautify or smooth the face. Change only the clothes and the background. Everything else stays the same.
Three findings you can use in ten seconds:
- The winning wording scored highest, but the margin is one photo. W4 scored 35 of 36 on our trait scale; the longest wording, with a full do-not list, scored 30 — and the entire gap comes from a single wedding photo (11 vs 6; the two wordings tied in the other two scenes). With three photos per wording, treat the ranking as a hint, not proof.
- No face instruction passed our “kept” test (3 of 3) — and still drifted more. The no-lock photos collected 7 “Close” marks out of 18 traits, against the winner’s 1. Gemini anchors on your uploaded photo, but the face line is doing real work: it is the difference between “mostly the same” and “the same”.
- The scene hurt more than the wording. The only trait that ever fully changed was hair: 3 times, all in the wedding scene, where formal styling made it darker and flatter.
Copy this
Paste this before any Gemini photo-edit prompt:
Keep the exact same face as the uploaded photo: same eye shape and spacing, nose width, lip shape, jaw, cheekbones, skin tone, and age. Do not beautify or smooth the face. Change only the clothes and the background. Everything else stays the same.
Note: if your own prompt also changes the pose or expression, drop the “Change only the clothes and the background” line — or add “except the pose and expression the prompt asks for”. That line is written for outfit-andbackground swaps, and it fights a prompt that asks for a new pose.
What we tested
One reference portrait (the author’s own photo), uploaded fresh for every edit:

- Test A (15 photos): 5 face wordings × 3 scenes — wedding stage, roadside chai stall, Hunza river. A new chat for every photo.
- Test B (6 photos): the winning wording, two ways — 3 edits chained inside one chat, each feeding on the last, vs 3 fresh chats.
- Validation (5 photos): the winning wording pasted in front of our 5 most-copied prompts, exactly as published on PromptTick.
The five wordings:
- W1 — No lock: scene words only, no face instruction at all.
- W2 — Short lock: “Keep the same face.”
- W3 — Detailed lock: the trait-by-trait line (eyes, nose, lips, jaw, skin tone, age; no beautifying).
- W4 — Detailed + one change: W3 plus “Change only the clothes and the background. Everything else stays the same.”
- W5 — Detailed + do-not list: W4 plus a list of everything not to change.
Results: Test A
Every photo was scored trait by trait against the reference: eyes, nose, lips, jaw, skin tone, and hair/age, each marked Same, Close, or Changed. A photo counts as “Kept” when at least 5 of 6 traits are Same or Close and neither the eyes nor the jaw Changed.
| Wording | Face kept | Trait points (of 36) | Same / Close / Changed (of 18 traits) |
|---|---|---|---|
| W1 No lock | 3 of 3 | 29 | 11 / 7 / 0 |
| W2 Short lock | 3 of 3 | 29 | 12 / 5 / 1 |
| W3 Detailed lock | 3 of 3 | 30 | 13 / 4 / 1 |
| W4 Detailed + one change | 3 of 3 | 35 | 17 / 1 / 0 |
| W5 Detailed + do-not list | 3 of 3 | 30 | 13 / 4 / 1 |
Three results worth a close look in the photos below:
Best of the test: chai stall with W4, a perfect 12 of 12.

Worst of the test: wedding with W5, the longest wording, 6 of 12. The face held, but the hair came back darker and flatter and the likeness sat slightly off.

Most surprising: wedding with W1 — a perfect 12 of 12 with no face instruction at all.

All 15 Test A photos, exactly as Gemini returned them — the weak ones included:















Results: Test B — same chat vs fresh chat
A common fear is that editing an edit degrades your face step by step. Over three chained edits — the length of this run, on one face — it did not:
| Method | Face kept | Trait points (of 36) |
|---|---|---|
| Chained in one chat | 3 of 3 | 36 |
| Fresh chat each time | 3 of 3 | 36 |
Chaining carried over things, not identity: the third chained photo (Hunza) still holds the tea cup from the chai-stall edit before it, and the collar style of the first render followed through the chain. If you chain edits, watch for leftover props. Your face is not the part that drifts.






Results: the winner on our 5 most-copied prompts
Finally, we pasted the winning wording in front of the five prompts PromptTick readers copy most, word for word as published:
| Prompt | Copies | Face kept? | Trait points (of 12) |
|---|---|---|---|
| White Shalwar Kameez at Faisal Mosque | 75 | Yes | 12 |
| Sky Blue Shalwar Kameez by the Hunza River | 67 | Yes | 10* |
| Nikkah / Formal Ceremony | 58 | Yes | 11 |
| Winter Shawl Portrait | 56 | Yes | 12 |
| Islamabad — Cream Shalwar Kameez | 54 | Yes | 12 |
*The Hunza River prompt itself asks for a smiling look-away profile, so the eyes and lips were marked Close for the camera angle, not for drift.
Worth knowing: the Faisal Mosque and Hunza River prompts contain no face instruction of their own. Adding the winning line is the whole face protection they get, and it was enough for a perfect and a near-perfect score.





The control check: can this scoring fail a face?
A fair question — put to us about this very test — is whether a “kept” verdict means anything if the test never fails anyone. So we ran three more edits designed to break the face on purpose, and scored them with the same sheet:
| Control edit | Verdict | Trait points (of 12) | What happened |
|---|---|---|---|
| “Make him 60 years old” (chai scene) | Drifted | 3 | Grey hair, deep wrinkles, sagging jaw — aged hard, exactly as instructed |
| “Replace him with a completely different man” (chai scene) | Drifted | 1 | Bald, glasses, round face, clean-shaven — a different person, as instructed |
| “Turn him into a glamorous fashion model” (wedding scene) | Kept | 9 | Gemini resisted the “different, more handsome face” instruction and retouched instead: airbrushed skin, sharper-looking jaw, styled hair |
Two honest conclusions. First, the scoring does catch deliberate identity breaks: an aged face and a swapped person both fail, hard. Second, the “Kept” bar is generous — a heavy beautify still passes it, and only the trait points (9, against the winners’ 12) show the cost. That is why this page publishes trait points and full trait counts, not verdicts alone.



Read nextPromptTick AI Photo Editing Benchmark: Gemini vs ChatGPT (Edition 1)10 min readHow we scored
- Date: November 1, 2026. Every photo was made that day, first try, with no re-rolls and no best-of selection. The weak ones are published above too.
- Tool: the Gemini app, signed in with a Pro account. The app’s model picker showed the label “Flash-Lite” during the test; we record the label exactly as shown.
- Reference: one portrait of the author, uploaded fresh for every non-chained edit.
- Scoring: one scorer (Muse, an AI vision model), working from a shuffled pack with the wordings hidden, marking 6 traits per photo against the reference, plus checks for hands, scene, outfit, and garbled text. Solo scoring is a limitation: read the verdicts as careful, not laboratory-grade.
- Sample size: 26 photos, one face, one day. This shows what happened in our test. It does not prove what Gemini does with every face.
- The “Kept” bar: generous by design — it asks “is this still recognisably the same person?”, and a heavy retouch can pass it (see the control check above). That is why every photo’s trait points are published alongside its verdict.
Download the complete results as a CSV — 29 rows, every trait mark for every photo, controls included.
Cite this test
“PromptTick tested 5 face-lock wordings on one person’s portrait across 26 first-try Gemini edits (Nov 1, 2026): all 26 passed the study’s face-kept rule, and a detailed lock ending in ‘change only the clothes and the background’ scored highest (35/36). A small test — one face, one day — so treat the ranking as a hint, not proof.” — prompttick.com/gemini-face-consistency-test/
Questions people ask
Why does Gemini change my face in some edits?
In our test the face itself barely moved. What changed was hair (3 times, all under formal wedding styling) and small amounts of smoothing. Heavy styling scenes plus long warning-list prompts were the weakest combination, not the lack of a face line.
Does writing “keep the same face” actually help?
Yes, in the detail. The short version scored the same total as no instruction at all (29 of 36), but look closer: the no-instruction photos collected 7 “Close” marks out of 18 traits, while the winning wording collected 1. Naming the traits and adding “change only the clothes and the background” scored best (35 of 36). Past that, more words stopped helping.
Should I edit in the same chat or start a new one?
For your face, over the three chained edits we ran, it made no difference (36 of 36 both ways). For cleanliness, fresh chats are safer: chained edits carried a tea cup from one scene into the next.
Do longer prompts keep my face better?
No. The longest wording in our test (W5, with a full do-not list) tied for second and produced the single weakest photo.
Try the prompts yourself
- White Shalwar Kameez at Faisal Mosque
- Sky Blue Shalwar Kameez by the Hunza River
- Nikkah / Formal Ceremony
- Winter Shawl Portrait
- Islamabad — Cream Shalwar Kameez
Related research
We also tested whether longer prompts edit better than short ones. In 27 blind-scored first-try Gemini edits, short ~30-word prompts outscored ~200-word prompts — read the Prompt Length Test.