You ask Gemini to "make me stand with my hand on my hip". The pose is great. The person is not you.
That happens because a pose change makes the model redraw most of the image, and the face gets redrawn along with the body. Ask for a new outfit and a new background in the same message and you've basically asked for a new photo of a new person.
Below is the order I use to keep the face: what stays, then what moves, one change at a time. The prompts come from real pages in the Promptclay AI pose prompts collection, which is free to copy.
Start with a photo that's easy to copy
Gemini can only hold on to what it can see. A sharp selfie or waist-up shot beats a full-body photo where your face is a few pixels wide. Front or three-quarter angles work better than a profile. Move hair, hands and sunglasses off your face if you can. And if there are two people in the photo, crop to one.
Put the face lock first
Order matters. A face rule tacked on after a long pose description is easy for the model to underweight. I open with the lock and close with a single pose line:
Using my uploaded photo, keep the same person: same face, same facial features, same hairstyle and hair colour, same skin tone, same body proportions, same outfit, same background and same lighting. Do not change the face. The only change is the pose: [new pose]. This is my own template, based on the pattern the better pose prompts on Promptclay follow.
Describe the pose in body parts
"Confident pose" makes Gemini guess. Guessing is where elbows bend wrong and faces wander. Write it like stage directions: weight on which foot, torso angle, a job for each hand, head and eye direction, and the crop.
Here's a real one that does it well. It describes the squat, gives each hand an object, then spends a whole sentence on the face:
Example image from the Street Style Portrait with Identity Lock and Specific Pose page (Nano Banana Pro). The original reference photo isn't shown on the page, so this shows the pose, not a verified face match.
Create an ultra-realistic photo of the uploaded person in a low squatting pose on an outdoor asphalt surface. The person is wearing an oversized white t-shirt, sheer black tights, and pointed beige high heels. One hand is holding a clear plastic cup with iced coffee and a straw, while the other hand holds a black smartphone. Dark nail polish and a silver ring on the fingers are clearly visible. The hairstyle, hair length, and hair color must remain exactly the same, with soft waves and a defined front curl. The face, identity, and all facial features must stay completely unchanged. The person is wearing black oval sunglasses and small earrings. Natural daylight, realistic shadows, street-style aesthetic, sharp focus, high detail, cinematic yet casual mood, no changes to clothing, accessories, pose, or proportions. One gotcha: if you aren't wearing sunglasses in your photo, delete that bit. Otherwise Gemini adds them and covers the part of your face it most needs to match.
Say what's allowed to change
My favourite pattern on the site is from the Consistent Series Character Street Snapshot prompt. It's written for GPT Image 2, but the structure works in Gemini too:
ABSOLUTE SERIES LOCK: Keep EXACTLY the SAME woman and identity, SAME city street, SAME enormous messy bouquet of {flower_type}, SAME {outfit}, SAME long loose hair, SAME {vintage_tint}, SAME heavy motion blur, SAME softness and SAME old RAW 35mm character. DO NOT redesign anything. ONLY CHANGE: POSE + CAMERA ANGLE. The rest of that prompt describes one pose in a lot of detail. You don't have to go that far. The useful bit is the explicit "ONLY CHANGE" line. Pose and camera angle together are usually fine. Pose plus outfit plus location is where faces go.
No words for the pose? Use a second image
Upload your photo, then a second image showing the pose (a mannequin or 3D figure is ideal). The Prompt for Changing Character Pose using Reference Image page has the shortest prompt in this post:
Change the character in the first image to the pose in the second image I'd add the lock so nothing else leaks over from image two:
Change the person in the first image to the pose in the second image. Keep the face, hair, outfit and background from the first image exactly the same. Take only the pose from the second image. If your pose reference shows a real face, Gemini may blend the two. Faceless references are cleaner.
Fixing drift
Same chat, one fix per message:
- "Keep this exact pose. Redo only the face to match my original photo."
- "Smaller change: only raise the right arm."
If it keeps failing, your new pose is too far from the original. Sitting to jumping, or front to back, means Gemini invents too much. Do it in two hops.
More pose prompts, each shown with its example image, are in the AI pose prompts collection on Promptclay.
