Video generation
A video is many images that must agree with each other over time. Learn why that is much harder and costlier than one image, and where today's models still slip.
In 60 seconds
Video generation
A video is many images that must agree with each other over time. Learn why that is much harder and costlier than one image, and where today's models still slip.
How video models work, in general
- 1
Prompt
Text, and sometimes a starting image: "A boat on the backwaters of Kerala at sunrise" - 2
Noise block
Start with random noise shaped like a short compressed video. - 3
Denoise together
Clean all frames at once, step by step, so they stay in agreement. - 4
Decode
Expand the latent back into full frames. - 5
Extend (optional)
Some tools make longer videos by generating the next part from the end of the last.
Why video is harder than images
| Challenge | What goes wrong |
|---|---|
| Consistency over time | A shirt changes colour, a person's face shifts, an extra finger appears for a few frames |
| Physics and cause | A glass breaks before it is hit; water flows uphill; a car passes through a wall |
| Object permanence | Something goes behind a tree and does not come out, or comes out as something else |
| Length | Errors build up over longer clips, so many tools make short clips |
| Data | Good video with accurate descriptions is harder to collect than captioned photos |
The cost of moving pictures
Is it a world simulator?
- Judging it by the first second
- Assuming a real-looking clip was filmed
- Sharing a shocking clip straight away
- Watch edges, hands, text and objects passing behind others
- Look for source, context and content credentials
- Check with trusted news before forwarding
Watch and read more
Lab
A frame-by-frame consistency check of one clearly labelled AI video.
The problem
Find a short video that is clearly labelled as AI-generated, for example on a video platform or a tool's own showcase page. Do not use videos of real people in private situations.
- Write the source and the label that says it is AI-made.
- Pick three things to track: a person or animal, an object, and a background feature (a sign, a building, a tree).
- Watch the video, pausing every second or two. For each tracked thing, note any change that should not happen: shape, colour, count, text, or position.
- Note one moment where physics or cause and effect looks wrong, if any.
- Write two sentences: would a casual viewer on a small phone screen notice these errors? What else would help them know it is AI-made?
You are done when
Hard questions
Try to answer before you reveal. If you can answer these, you understood the lesson.
Q1Why does generating all frames together (rather than one by one) help consistency?Reveal
If each frame is made separately, small random choices differ each time, and nothing forces them to agree. When all frames are denoised together, attention lets each part of each frame look at the others at every step, so they settle on the same face, colour and layout together. The cost is much more computation and memory, which limits length.
Q2A news team wants to use AI video to "reconstruct" a real event for viewers. What are the risks and what would you require?Reveal
Viewers may take it as real footage, and the model may invent details that did not happen. Require a clear, visible label for the whole time it is on screen, a script checked against verified facts, no realistic depiction of real identifiable people without consent, content credentials on the file, and a record of how it was made. Consider whether a simple diagram would inform people better with less risk.
Questions people ask
Why do many AI videos last only a few seconds?
Because cost grows with length, and errors in consistency build up over time. Making longer videos usually means stitching or extending shorter parts, where joins can show.
Can a video model keep a character the same across scenes?
It is a known hard problem. Some tools let you give a reference image of the character to help. Even so, small details like jewellery, logos or tattoos often drift. Check each shot.
Is detecting AI video easy because of these errors?
Less and less. The errors are getting rarer and harder to spot. Visual clues help, but source checks, context and content credentials are more reliable than your eyes alone.
What good uses does video generation have?
Storyboards and drafts for filmmakers, explainer animations for teaching, product mock-ups, and small businesses making simple adverts. Being clear with viewers that it is AI-made keeps trust.
Lesson test
5 questions. Get 3 right (60%) to pass and complete this lesson.