Why AI B-Roll Looks Cheap — and How Theme-Locking Fixes It

Watch almost any video assembled from independently generated visuals and you will feel something is off before you can say what. The individual frames are usually fine — often better than the stock footage they replaced. The problem is that frame 3 and frame 17 come from different worlds.

Humans are extremely good at detecting inconsistency in visual language and extremely bad at articulating it. Viewers do not think "the colour temperature shifted and the illustration style changed from flat vector to painterly". They think "this looks cheap" and scroll.

Why per-scene generation drifts

If each scene is generated from its own prompt, every generation independently re-decides everything the prompt did not pin down: lighting, palette, level of abstraction, camera distance, rendering style, how figures are drawn. Those defaults are effectively random per call.

Even careful prompting only partially helps, because the drift is in the unspecified dimensions and there are more of them than you can enumerate per scene. Twenty scenes means twenty independent rolls, and inconsistency compounds visibly.

What a theme actually is

A theme is a fixed decision set applied to every generation in a project:

• Palette — a small committed set of colours, including what backgrounds do
• Rendering style — flat vector, painterly, photographic, diagrammatic, collage
• Level of abstraction — literal depictions versus conceptual imagery
• Camera language — distance, angle, whether shots are static or move
• Type treatment — where text appears, in what face, at what weight
• Subject treatment — how people, objects, and data are drawn, consistently

Locked once at the top of the project, injected into every generation, unchanged for the whole video. The individual frames may be no better than before; the video will look dramatically more expensive, because the eye reads consistency as intent.

Consistency compounds across uploads

The same argument applies one level up. A channel where every video has its own aesthetic is a channel with no visual identity, which means every thumbnail competes from zero.

Channels that grow fast almost always have a look you can identify at thumbnail size with the title cropped out. That recognisability is a click-through advantage on every future upload — and it costs nothing except the discipline to keep the theme fixed.

The four failure modes worth checking for

Style drift. Scene 3 is flat illustration, scene 11 is photoreal. Fix: one rendering style per project, no exceptions for "this one scene needs a photo".

Palette drift. Warm ambers early, cold blues later, no narrative reason. Fix: a committed palette, with deviation reserved for moments where the script changes register.

Density mismatch. Some frames are busy and detailed, others nearly empty. This reads as sloppy rather than dynamic. Fix: pick a density band and hold it.

Literalism. Every noun in the narration gets a matching picture. A script that says "the company collapsed" does not need a picture of a collapsing building — visuals that merely restate the narration add nothing and make the video feel like a slideshow. Fix: visuals should extend the point, contrast with it, or provide evidence for it.

Motion: less than you think

The instinct with generated video is to make everything move. Constant motion in b-roll competes with the narration for attention, and narration is what carries an explainer.

A workable rule: motion should be slow enough that you can read a sentence over it without losing either. Slow pushes, gentle parallax, and long holds outperform fast pans and rapid cuts in long-form, where the viewer is listening rather than watching.

The audit that takes two minutes

Before you export, scrub the timeline at high speed with the audio off and watch only for whether it looks like one video. You are not evaluating individual frames — you are looking for the moment your eye snags. Every snag is a consistency break, and fixing four of them is usually the difference between "AI-generated" and "well-produced".

Then watch it once at normal speed with audio, asking a single question per scene: does this visual earn its place, or is it just something on screen? Anything that fails becomes a longer hold on the previous shot. An empty frame that matches the theme beats a busy one that does not.