If you make AI video from drawn frames, you have probably seen this. You plan a short reel where a character walks down a street, stops at a corner and walks into a café. You draw the frames. Every one of them shows her in the middle of the picture, the same shop window behind her, from the same distance. She is "walking", but she never gets anywhere.
We ran into exactly this while testing ads for our own AI influencers. The face, the outfit and the product all stayed consistent, which we had worked hard for. The problem was that the consistency had spread to everything else, including the parts that should change.
This article explains why it happens and what fixed it, step by step, from our own side-by-side tests.
A storyboard should change angle, distance and place from card to card.
How a storyboard frame is usually made
Most image-to-video workflows look like this:
Draw a set of still frames, one for each shot or beat.
Check them, because a still costs cents and a video costs dollars.
Give the stills to a video model as start, end or keyframe images and let it animate between them.
To keep the character consistent across the frames, every frame gets reference images: photos of the character's face, and often an earlier frame of the same scene. That earlier frame is what keeps the room, the light and the outfit the same.
The trouble is in how that earlier frame is described to the model.
A good AI UGC ad is decided before anything is generated: in about 60 words of script. Here is the five-beat structure our AI influencers' producer writes to, how many words fit in 20 to 30 seconds, and the mistakes that make a script sound like an ad.
In our case, every frame after the first was given the first frame with an instruction along the lines of "this is the same place: keep the background, light and time of day exactly as in this image". For a scene at one café table, that works well. For anything where the character moves, it is the wrong instruction.
The image model does what it is told. It copies the first frame's background, and with it the first frame's camera distance and framing. So a planned close-up comes back as a medium shot, and a planned shot "thirty metres further down the street" comes back in front of the same bakery window. The better the original plan was at varying shots, the more of that variety got flattened.
There was a second cause, which matters just as much: nobody had decided where the camera was. The shot descriptions said what she did ("she stops at the kerb and looks at her hands"), but not where she was on the route, where the camera stood, which way it faced or what was behind her. Without that, image models fall back to their favourite composition: the person centred and facing the camera, the street disappearing behind them.
Fix 1: block every shot before you describe it
Film crews call this blocking: deciding where the actors are and where the camera goes before anything is shot. For AI storyboards it is the single biggest fix we found. For each frame, write three things down first:
Where: her exact spot, and how far she has moved since the last shot. "Thirty metres further down the street, at the bakery's door."
Camera: where it stands, its height, its distance and which way it faces. "Behind her right shoulder at eye height, two metres back, facing north toward the tram stop."
Behind: what is behind her in this shot and in no other. "The yellow tram and the crossing."
Then start the frame's prompt with the camera and the background, before what she does. In our tests, adding this blocking was the difference between five frames of the same spot and a sequence that actually travelled from a street corner past a bakery to a café door.
A few rules of thumb help:
No two frames share a camera position and background unless the character is standing still in the same spot.
Vary the angle the way an editor would cut it: front, side, from behind, over the shoulder, low, close.
Make close-ups explicit. "A close-up at chest height, 40 cm away, filling the frame with the product on her jacket" survives far better than "a closer shot".
Even with blocking, image models drift back toward a front-on medium shot whenever a description is vague. The more concrete the camera line, the better the angle holds.
Decide where the camera stands for each shot, not just what happens.
Fix 2: give each reference picture one job
Reference images are powerful, and models take them literally. The fix is to tell the model exactly what each one is for, and what it is not for.
What worked best for us:
The character's reference sheet first. A sheet showing the face from several angles kept the face consistent on its own. A single portrait added on top tended to pull poses and framing toward that portrait.
The previous shot, not the first one. Each frame sees the shot just before it, so changes build up naturally. The instruction says to keep her face, her outfit, the day, the light and the weather, and explicitly says that the framing and background come from this shot's description, not from the picture.
Product photos only where the product is in shot. When we attached a product's photos to every frame, its logo started turning up printed on a tote bag. Now product photos only go to frames that show the product, with an instruction that it appears once.
One more small one: say that bags, cups and clothes are plain, with no logos or lettering. Image models like to invent brand names, and one of our test frames came back with a real fashion label printed on a tote bag.
Fix 3: check the frames as a set, not one by one
Our first review step looked at each frame on its own and asked "is this a good, realistic photo?" Most of the static frames passed, because each one was a perfectly good photo. The problem only shows when you look at them together.
So the check now happens on the whole set, side by side with the plan:
Does each frame show what its shot is meant to show? A door being pushed open, not a door already open.
Is the product there wherever the shot needs it, and is it the real one, checked against its photos?
Does the background move on when the character moves?
Does the plan's timing leave room for each action and line, and does the last shot end on a settled moment?
Is a payoff, like a final look to camera, saved for the end and not spent in an earlier frame?
The reviewer then decides: go, redraw specific frames, or send the plan back to be rewritten. A whole-set check on our tests caught problems the frame-by-frame check passed, including two frames that were almost duplicates and a key action that never happened.
A checklist for your next AI storyboard
Write the blocking first: where, camera, behind, for every frame
Start each frame's prompt with the camera and the background
One reference sheet for the face; drop extra portraits that pull poses
Show each frame the shot before it, for face, outfit and light only
Attach product photos only to frames that show the product
Keep props plain, with no logos or lettering
Review the frames as a set against the plan before making any video
Every Influgen influencer has a team that plans posts like a small production crew. Her creative director blocks each shot, every frame is drawn from her reference sheet and the shot before it, product photos only go where the product appears, and her critic reviews the whole set against the script before you see it. You approve the frames before any video is made.
Most AI ads give themselves away through the sound before the picture: dead air, a voice with no room around it, music fighting the words. Here are the seven fixes we now build into every talking ad, with the exact settings we use.
Over nine days, our review step made 672 checks of AI-generated frames for 124 posts and turned down more than half. The surprise: bad hands were rare. Most frames failed because they didn't show what the post needed.
