Which AI video generator gets physics right?
Winner: Kling 3.0, averaging 7.2/10 across four physics-heavy prompts. Nobody nailed the kickflip, which shows where all four models still break.
| Prompt | Kling 3.0 | PixVerse V6 | Runway Gen-4.5 | Veo 3.1 Fast |
|---|---|---|---|---|
| Dog catches frisbee | 7.8 | 7.8 | 6.0 | 5.8 |
| Kickflip down stairs | 5.5 | 6.0 | 6.8 | 6.0 |
| Origami crane takes flight | 7.5 | 6.8 | 6.3 | 6.8 |
| Milk into coffee | 8.0 | 8.0 | 8.3 | 6.5 |
| Average | 7.2 | 7.1 | 6.8 | 6.3 |
Dog catches frisbee
“A golden retriever catching a frisbee mid-air on a beach at sunset, slow motion.”
adherence 9motion 8quality 7artifacts 7
Frisbee flies in, dog leaps and catches it; slightly flat colour.
adherence 8motion 8quality 8artifacts 7
Dog leaps and catches the frisbee mid-air; warm beach light but not a clear sunset.
adherence 5motion 6quality 7artifacts 6
Dog already has the frisbee; no mid-air catch; ends sitting.
adherence 5motion 6quality 6artifacts 6
Backlit dog running with frisbee in mouth; no catch shown.
Kickflip down stairs
“Skateboarder doing a kickflip down stairs, tracking shot.”
adherence 7motion 6quality 8artifacts 6
Board flips, landing on flat ground; stairs barely part of the trick.
adherence 6motion 5quality 7artifacts 6
Board flips over stairs but skater stays at the top; physics off.
adherence 6motion 5quality 7artifacts 6
Board flips off a ledge, but the skater never really goes down the stairs; landing physics off.
adherence 4motion 5quality 7artifacts 6
Rides a ledge; no kickflip, no stair descent.
Origami crane takes flight
“Paper origami crane unfolding itself and flying away, stop-motion style.”
adherence 8motion 7quality 8artifacts 7
Unfolds from flat into a crane and takes off; stop-motion feel.
adherence 5motion 6quality 8artifacts 8
Clean finished crane that lifts off; no unfolding, no stop-motion look.
adherence 5motion 6quality 8artifacts 8
Clean finished crane that lifts off; no unfolding, no stop-motion look.
adherence 7motion 6quality 6artifacts 6
Paper rises and flies but folds look odd; grainy.
Milk into coffee
“A hand pouring milk into black coffee, the swirl in macro, top-down.”
adherence 8motion 8quality 9artifacts 8
Beautiful swirl in a mug; angled rather than top-down.
adherence 8motion 8quality 8artifacts 8
Near top-down glass, clean realistic swirl.
adherence 8motion 8quality 8artifacts 8
Realistic milk swirl in black coffee, near top-down; pitcher partly in frame.
adherence 7motion 6quality 7artifacts 6
Milk forms unrealistic blobs.
What this means for your prompts
- Describe the key moment, not just the scene. Veo and Runway both showed a dog with a frisbee but skipped the catch itself.
- Complex body physics (tricks, sports) still fail. Plan to generate several takes, or cut around the hardest moment.
- Transformations work better on Kling. Only Kling showed the origami crane folding itself before taking off.
Related: Veo vs Kling · Kling 3.0 review · PixVerse vs Kling · AI video prompt guide · Best for food · Most realistic AI video generator
FAQ
Which AI video generator has the most realistic physics?
In our four physics-heavy prompts (frisbee catch, kickflip, origami flight, milk pour), Kling 3.0 averaged 7.2, ahead of PixVerse V6 (7.1), Runway Gen-4.5 (6.8) and Veo 3.1 Fast (6.3). Kling and PixVerse V6 were the only models that showed the dog actually catching the frisbee mid-air.
Can AI video do a skateboard kickflip?
Not convincingly yet. The kickflip down stairs was our hardest prompt: the best score was 6.8. Runway flipped the board but landed on flat ground, Veo flipped the board while the skater stayed at the top, PixVerse flipped the board off a ledge but never went down the stairs, and Kling skipped the trick.
Which is best for liquids?
Runway Gen-4.5 (8.3) Kling 3.0 (8.0) and PixVerse V6 (8.0) all made a realistic milk-into-coffee swirl. Veo's milk formed unrealistic blobs (6.5).
How were these scored?
Each clip was scored blind from 1 to 10 on prompt adherence, motion, visual quality and artifacts, and the overall score is the average. It's one 4-second 720p generation per model per prompt.