One face for the whole video
Lesson 13 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
Why the face drifts
We’re starting the character level. For the model, every generation is a new world: it doesn’t remember what Mira looked like in the last shot. If the prompt only says “a 19-year-old girl in a hoodie”, the model draws a new girl every time. Even with identical text, small things drift: face shape, hair length, hoodie colour, age, earrings. It’s one of the best-known weaknesses of video models: reviews of Kling 3.0, for example, name fingers and face drift as its main minuses. For a brand that’s critical: viewers should recognise your hero from the first second of every video.
The character sheet
Animators and game studios have made model sheets for decades so dozens of artists can draw the same hero. Do the same. In an image model, generate Mira from six sides: front, three-quarters, profile, full body and two emotions, say laughing and a serious face. Important: the same clothes, even soft light and a plain background, so the model sees the person, not the surroundings. Pick the best version and don’t change it again. This sheet becomes the main reference for every shot, and the face stops “wandering”.
The character passport
The image holds the face; the text holds everything else. Make a character passport: a short block you paste into every prompt word for word. Who: name, age, a couple of features. Hair: length, colour, bangs. Clothes: colour, material, cut. Details: a coral beanie, one earring in the left ear and no other jewellery. Why word for word: change “coral” to “salmon” and the model will dutifully recolour the hoodie. Pros keep this block in their notes and copy it rather than rewriting it. Look at Mira’s passport in the diagram: four lines that cover almost all the drift.
References in the model
Modern models accept reference images right in the generation. Veo 3.1 takes up to three; Seedance 2.5 in Higgsfield Cinema Studio takes up to fifty. More isn’t better: each image should be responsible for one thing. The face sheet holds the look. The clothing on its own on a plain background holds the hoodie: colour, print, fit. The location holds the place if the video has several shots in one spot. Don’t mix a face, different clothes and someone else’s background in one image: the model will take a bit of everything. And the same set of references should go into every shot of the video.
Logos and lettering
Letters are a weak spot of AI video. Even if the logo comes out clean in the first frame, letters start to wander as the fabric and camera move. For a brand that’s unacceptable: a wonky logo looks fake. Two working paths. First, give the model a photo of the real hoodie as a reference, so the print comes from the actual item instead of being invented. Second, the most reliable: generate the hoodie without the logo and add the logo in the edit with tracking, where the software sticks it to the fabric and follows the motion. And avoid close-ups of lettering in motion: the more letters on screen, the higher the chance of a glitch.
The grid check
When all the shots are ready, do a simple trick: take one still from each shot and lay them out side by side in a grid. Individually, the eye forgives small drift, but side by side you instantly see that shot five has a different beanie and hair twice as long. Check point by point: face, hair, clothes, logo, accessories, light. Redo the odd shot with the same references and the same passport instead of “adjusting” the whole video to it. It takes five minutes and saves you from comments like “why is she different in every frame?”.
The character folder
Everything from this lesson goes into one folder, and it works for you on every future drop. The character sheet, the text passport, the palette and light from lesson 6, the clothing photos. For a new video you don’t start from zero: you take the folder and write only the action and place, while the look, clothing and style are already locked. At the end, the grid check. That’s how brands with recurring characters work: the hero is recognisable, and each new video is faster and cheaper to make. Mira built this folder once and now makes a drop video in an evening.
Check yourself
- Across the ad’s five shots, Mira has different hair lengths. What helps most reliably?
Correct answer: The same text passport and the same references in every shot. The model doesn’t remember previous shots. Identical text word for word and identical images hold the look. - The brand logo on the hoodie turns into scribbles in motion. What should she do?
Correct answer: Generate the hoodie without the logo and add it in the edit with tracking. Letters in motion are a weak spot for models. A real logo stuck on with tracking is always clean. - Mira is making a character sheet. Which background and light should she choose?
Correct answer: A plain background and even soft light. The sheet must show the person. A neutral background and light stop the model from carrying surroundings and colour casts into every shot. - In the grid check, shot 4 has a slightly different face; the other five match. What should she do?
Correct answer: Redo shot 4 with the same references and passport. Fix the one that stands out. There’s no need to touch the five good shots.
← Motion control and the weight test Cutting without jumps →