Video and its current limits

Video is the fastest-moving area in this module, which makes anything written about it age quickly. So this lesson is about the shape of the problem, which changes far more slowly than the capability.

Read the specifics here as an example, and check the current state yourself before relying on it.

Consistency across time is the hard part

An image has to be coherent in space. A video has to be coherent in space and across time — the coat stays the same colour, the room keeps the same furniture, the face remains the same face.

Nothing in the generation process guarantees that. So the characteristic failures are drift and morphing: details that change between frames, objects that quietly become other objects, a background that reorganises itself when the camera moves.

Longer clips are worse, because there is more time to drift. That single fact determines how you should work.

Where it is genuinely useful now

Honestly assessed, and biased towards what holds up:

• Short atmospheric b-roll. Clouds, water, a street at night, abstract texture. No people, no text, no dialogue. This works well.
• Simple camera moves over a scene — a slow push, a drift.
• Animating a still image you already have and like, for a few seconds.

Where it still falls apart

• Anything longer than a few seconds without a visible cut.
• Hands and faces in motion, especially speech.
• Text, for all of L1’s reasons plus the requirement that it stay stable.
• Physical cause and effect — pouring, cutting, breaking. The result often looks correct frame by frame and impossible as a sequence.
• A specific character across multiple shots.

Work in shots, not films

The practical consequence: do not ask for a finished sequence. Plan like an editor.

Write a shot list — four seconds each, one idea per shot. Generate each separately, several times. Assemble in an editor.

This is how the work survives contact with the limitations. Cuts hide drift. Short clips are cheap to reroll. And a sequence assembled from deliberate shots has a structure, where a generated “video about X” has none — the same argument E5·L6 made about slides.

The disclosure question arrives here

Video crosses a line images mostly do not. A still can be obviously stylised; a few seconds of realistic video of a place or a person reads as footage, and footage is treated as evidence.

That makes labelling more important here than anywhere else in the module, and L7 deals with it properly. For now: if it could be mistaken for a recording of something real, it needs to say what it is.

Try it now (6 minutes)

Write a four-shot list for a fifteen-second sequence. One idea per shot, no people, no text.

Generate the easiest shot. Then generate the hardest and compare — the gap between them is the current state of the technology, measured by you rather than by a marketing page.

Check your understanding

1. Temporal consistency means:

Clips rendering at a steady frame rate Objects staying the same from frame to frame Generation finishing in predictable time

2. Why plan a shot list rather than request a finished sequence?

Shot lists are cheaper to generate Consistency degrades with length, and cuts hide drift Models cannot produce more than one shot

3. Which is most reliable with current video models?

A person speaking to camera Short atmospheric b-roll with no people or text A character appearing across several shots

Check answers

Recap

Video must hold together across time as well as space, and nothing guarantees it, so drift grows with length. Work in four-second shots, assemble in an editor, and stay with atmospheric material rather than faces and speech. And because video reads as footage, anything that could pass for a recording needs to say what it is.