How to keep an AI persona consistent across a batch
Updated October 5, 2026 · By Personifex
This is how the operator who builds Personifex keeps their own personas consistent. The face comes from reference photos, never from words: the prompt changes the scene, the outfit, the pose and the light, and a review pass catches what drifts.
No model keeps a face identical in every image, so the method is a batch and a review, not one perfect shot.
Four reference photos
Give each persona four photos that show the face from every side a model will need:
- A close-up of the face.
- The face turned to the left.
- The face turned to the right.
- The full body, head to toe.
Vary the light and the expression across the four, and shoot them on a plain white background, so the model learns the person rather than one room or one smile.
In Personifex a persona holds up to 5, 20 or 50 reference photos depending on the plan, and each run sends up to 4 of them. A good image already in your Gallery can become one without a download: right-click it and choose Add to library.
All four, or one
It depends on the model and on you. Send all four so the model knows exactly what the persona looks like, or one when the model is strong enough to hold the face from a single photo.
Models that hold a face
When the face matters, this operator trusts three: Seedream v5.0 Pro, GPT Image 2 and Nano-Banana Pro.
On GPT Image 2, set Quality to Medium or High, never Low.
Each model's prices are on its page, and the cheapest model for each job compares them.
Never describe the face
The prompt never describes the face, the hair, the skin or the age. Those come from the reference photos, and words about them compete with the photos. Describe what changes from shot to shot: the scene, the outfit, the pose and the light.
Ask for a phone photo. Phrases like “grainy iPhone 15 Pro look” and “not professionally taken” keep the result looking like a picture someone took. Leave out “perfect lighting” and “studio lighting”: a phone photo has neither.
The aspect ratio makes little difference to consistency.
Judge the batch
Rate every result in Review with the number keys. Two of the four ratings do most of the work:
| Rating | When | Where it goes |
|---|---|---|
| 1 Bad | It has the AI look, an extra finger or limb, it does not look like the persona, or it simply is not to your taste. | Trash, restorable |
| 2 Good | A keeper. | Gallery |
| 3 Great | The best of the batch. | Gallery |
| 4 Needs edit | One thing is wrong and the rest is a keeper: a finger, a wrong logo, odd text, something strange in the background. | Gallery, Needs edit |
Fix a near miss
Run the prompt again: in Prompts, Queue again puts a copy back in the queue for the next batch.
Or fix the one thing that is wrong: Edit with AI in the Gallery repaints just the part you mark. The result arrives in Review as a new image, and the original stays as it is.
What this cannot promise
No model keeps a face identical in every image. Good references, a model that holds faces and a prompt that leaves the face alone raise the share of keepers, but part of every batch drifts. That is why the method ends in a review pass, and why this guide gives no keeper rate: it depends on the persona, the model and your taste.