Nothing makes a faceless video look cheaper than a character whose face changes every scene, or a style that jumps from watercolour to photo and back. AI image tools don't remember the last picture they made. Every picture starts from nothing, so keeping a whole video consistent is your job, not the tool's.
Here is what works for us on rethink normal, where one recurring character appears in video after video.
Why pictures drift
Each request to an image model is read on its own. If you describe the same person in slightly different words, you get a slightly different person. If you leave out the style, the model picks one. Small differences add up over forty scenes.
1. Write a style sheet, and use it word for word
Before the first picture, write a short paragraph that fixes the look: the kind of picture (flat illustration, painted, photo), the colours, the light, how thick the lines are, how close the camera is. Paste the same paragraph into every request. Don't rephrase it, even slightly. Consistency comes from repeating the exact words.
2. Give every character a reference
For each person who appears more than once, make one picture you are happy with and keep it as their reference. Write down the details that must never change: hair, skin tone, age, clothes, colours. Attach the reference picture to every new request where the tool allows it, and repeat the fixed details in words as well.
Test the reference early: make the same character in four different poses before you commit. If the face doesn't hold across four, it won't hold across forty.
3. Reuse instead of remaking
If a scene shows the same place or moment as an earlier one, use the earlier picture again. It is free, and it is perfectly consistent. Mark repeats when you plan the video, before any pictures are made. In one of our early videos, the plan asked for thirteen pictures of what was really the same scene. One would have done.
4. Change one thing at a time
When a picture is almost right, ask for one change: "make it evening", "she is smiling". Asking for three changes at once often brings back a different character or a different style. Keep the version you liked until the new one is clearly better.
5. Keep words out of the picture
Image models still misspell signs, labels and numbers. Ask for pictures without text and add any words in the edit, where they are sharp, correct and in your own font.
6. Look at them all together before you render
Lay every picture out side by side on one page, in order, before you put the video together. A face that drifted, a colour that doesn't belong, or a scene in the wrong style jumps out at once on a sheet, and hides when you look at pictures one by one.
7. Keep a record
Save the exact words used for every picture you keep. When you need the same character next month, you start from what worked instead of starting again.
Where vbe | video build engine fits
vbe is a Mac app being built around these habits: one look for the whole video, a character reference kept with the project, repeated scenes marked and reused when the scenes are planned, and the cost of new pictures shown before any are made. See how it works.
Written by Aman, founder of vbe | video build engine and maker of rethink normal. Published 24 October 2026.