Why AI Video Works in Short Shots

Video models produce impressive four-second clips. Ask for twenty seconds of continuous action and objects start to melt. This is not a defect you can prompt around — it is a consequence of how frames are predicted.

Where the errors come from

• Error accumulation. Each frame is generated with reference to previous ones, so small inaccuracies compound.
• Identity drift. Faces and clothing shift gradually across a long take.
• Physics approximation. Objects pass through each other; liquids behave unconvincingly.

Designing around it

• Build sequences from multiple short shots, the way conventional editing already works.
• Cut on movement so the transition hides the join.
• Use image-to-video with a strong starting frame to lock composition.
• Keep camera motion deliberate rather than letting the model improvise.

The practical result

Treated as a shot generator rather than a film generator, current models are genuinely useful for previsualisation, transitions and b-roll. Accepting that framing produces far better output than fighting it.