The best AI video generators, tested on real projects

I gave every major AI video generator the same three briefs: a slow push through an empty office at dawn, someone’s hands typing at a desk, and a coffee cup on a windowsill with rain outside. Boring on purpose. These are the shots I actually need when I’m editing, not the dragons and spaceships every demo reel is full of.

The results told me more about where this technology is useful than any showcase would. Two of the three briefs came back usable. The hands did not, and hands are still the thing that gives it away.

What AI video generators are actually good at

The honest boundary, from running a lot of these prompts:

They excel at: atmospheric shots, landscapes, weather, abstract motion, textures, light through windows, aerial-style movement, anything where nothing specific has to be correct. My empty office and my rainy window both came back genuinely good, and I’d have happily cut either into a video.

They fail at: hands doing tasks, faces holding an expression over time, text of any kind, recognisable products, and anything with continuity across shots. My typing hands had six fingers in one generation and melted into the keyboard in another.

That split has been stable for a while now, even as quality has improved. The models got much better at rendering; they didn’t get better at knowing how many fingers a hand has.

The clip length problem

The constraint nobody puts in the marketing: you get a few seconds per generation, and there’s no memory between generations.

That means you can’t ask for a character in shot one and get the same character in shot two. Lighting shifts, faces change, the room is subtly different. For narrative work this is disqualifying. For b-roll it doesn’t matter at all, because cutaways are two seconds and nobody expects continuity between a shot of rain and a shot of a desk.

Which is why my strong recommendation for most solo creators is to stop thinking about generating videos and start thinking about generating shots. Use them the way you’d use a stock library, dropped into an edit you built normally. That workflow is covered properly in the AI b-roll tools that fill the dead space in your edit, and it’s where these tools genuinely earn their subscription.

How to prompt for usable footage

Most disappointing generations are prompt problems. What consistently improves results:

• Describe the camera, not just the scene. “Slow dolly forward, shallow depth of field, 35mm” does more work than three sentences about mood.
• Name the light. “Late afternoon light through a window, warm, long shadows” is the single highest-impact addition you can make.
• Keep it to one subject and one action. Two things happening produces confusion in both.
• Avoid people doing tasks. If a person must appear, keep them distant, in silhouette, or out of focus.
• Generate four, keep one. The variance between generations of the same prompt is enormous. Anyone getting good results is generating a lot and discarding most.

That last point is the practical reality and it’s why credit limits matter more than any quality claim. A tool that produces beautiful output but gives you twelve generations a month is worse than a decent one that gives you two hundred.

Where the money goes

Every tool in this category meters generation, and the pricing is genuinely confusing on purpose. Things to compare before subscribing:

Generations per month, not “credits”, because credit costs vary by resolution and length. Work out the actual number of clips.

Resolution on your tier. Many tools reserve higher resolutions for expensive plans, and upscaled 720p in a 4K timeline looks exactly like what it is.

Commercial rights. Frequently absent on free tiers.

Queue times. On busy free tiers you can wait twenty minutes for a clip you’ll discard.

I’d budget for whichever plan gives you enough generations to throw away four out of five, because that’s the real hit rate.

The honest alternative

Before you subscribe to anything: your phone.

For a substantial share of the b-roll a solo creator needs, walking around your own workspace for ten minutes with a phone produces more useful, more authentic footage than any generator. Your desk, your keyboard, your window, your coffee. It costs nothing, it’s genuinely yours, and it doesn’t have the faint synthetic sheen that a growing number of viewers actively notice.

I use generated clips for shots I can’t film. Rain when it’s sunny, an empty office when I work from a bedroom, an aerial when I’m at ground level. That’s a real and useful category, and it’s much narrower than the marketing suggests.

For assembling all of it, CapCut handles quick vertical edits well, and my CapCut review covers where it fits. If you’re deciding between editors, CapCut vs Descript is the head-to-head.

Which should you use?

• You need occasional atmospheric b-roll: whichever tool’s free tier covers your volume. You won’t use it enough to justify a subscription.
• You publish daily short-form: pick on generations per month, not on demo quality.
• You need consistent characters or narrative: none of them, yet. Film it or use stock.
• You need product shots: film the product. This is not close.

My take

The gap between what these tools demo and what they deliver on a Tuesday afternoon is still wide, and it’s wide in a specific, predictable way: they’re great at atmosphere and bad at specifics.

Once I stopped fighting that and started using them only for shots where nothing has to be correct, they became genuinely useful and I stopped being disappointed. Rain on a window when I need rain on a window. An empty room. Light moving across a wall. Two seconds each, dropped into an edit, and nobody ever asks about them.

What I’d avoid is building a workflow that depends on generating your main footage. Every creator I’ve watched try that has ended up spending more time regenerating clips than they’d have spent filming, and the result still looked slightly wrong in a way they couldn’t fix. Generate the gaps, film the substance.