A Midjourney user described in a post archived May 28, 2026 that they dropped an earlier image into Midjourney and asked for the same person in another setting. The new images looked like different people. I’d shrink the next request to one change and decide which details make the character recognizable before generating again.
The short answer
Choose one clear picture of your fictional character and write down the details that make them recognizable. Use that picture as a reference if your tool supports it. Change one thing in the next scene, then compare the images side by side. If a signature detail changed too, use the original reference again and make a narrower request.
Pick an original fictional character and a picture you’re comfortable using as a reference. You can practice the comparison with images you already have before generating another.
A second Midjourney user posted on November 29, 2025 about faces that still changed after reusing a prompt and trying reference images. The two posts describe individual results. The lesson I’d take is to define what you’re checking and keep the first picture close by.
How to keep the same character in a new scene
Pick an image where the identifying details are visible
If you do not have a picture yet, generate one simple front-view portrait of your original character and choose a version with all identifying details visible. Use that clear picture as your anchor. For a practice character, imagine a little gray robot with a square head, a single round amber eye and a teal stripe on its left arm. The lamp behind the robot can change. Its face and stripe are the details you want to keep.
Write a keep list and one change
Put the character details on three short lines. On a separate line, name the scene change. Start with something small, like moving the robot from a workbench to a station platform while keeping the same front view. If you change the camera angle and the lighting at the same time, a missing stripe can be hard to interpret.
Use the reference feature your tool supports
If your tool can use a picture as a subject reference, give it the anchor image and the short request. If you only have text input, repeat the exact keep-list wording in every prompt, including a fixed phrase such as “square-headed gray robot, one amber eye, teal stripe on left arm.” Compare each result against the anchor even if the tool cannot accept an image. Midjourney’s Omni Reference documentation describes the feature as available only in version 7. That is a Midjourney-specific option, not a requirement for this exercise.
Review the two images beside each other
Match the head shape, eye and stripe separately. A new platform and the same three cues pass this exercise. A platform scene with two amber eyes needs revision even if the colors look right. If the arm is out of frame, mark the stripe ‘can’t see’ and try a view that shows it. Use the anchor again instead of treating a partly changed image as your new reference.
A prompt to try
Use the supplied image as the reference for an original fictional robot. Keep the square head, the single round amber eye and the teal stripe on its left arm. Change only the setting to a station platform. Keep the front view and show both arms.
We can work through the keep list and one test image together in a session. See how the one-to-one session works.
What I’d double-check
I’d keep the first accepted image somewhere easy to find. Replacing it with every new draft makes small changes harder to notice. A reference gives the tool a starting point, so check the details yourself each time. If the scene hides one of your cues, give it an angle that makes the check possible.
Explore help with AI image prompts
Try it on your own project
A short series gets harder when the camera moves and the character still needs to read as the same one. I’d test the angle first, then the scene. We can decide which cues work from both views and which need a different framing.
Work through it with me
Bring an original character to a 60-minute virtual session. We’ll choose an anchor image, write a keep list, and compare an angle change with a scene change. You’ll leave with a reference brief and a clear next edit for any image that still drifts.
Last checked November 7, 2026.