Best AI model for videos of people
Real-looking people doing something: dancing, cooking, running, laughing with friends.
Early result: 4 prompts per video task. See the all-video ranking for the stronger signal.
Best for videos of people
Wan 2.7
- 28%chance of ranking first
- Kling V3runner-up
Judge's most likely first pick; first on 2 of 4 requests.
Judge agreement 0.50
Tested requests, made with Wan 2.7
-
three friends laughing together at an outdoor cafe table, one tells a story with hand gestures, golden hour, candid lifestyle video
-
A young woman dancing freely in a sunlit living room, loose white shirt, joyful, handheld camera, natural colors, social media video
-
chef chopping fresh vegetables quickly on a wooden board in a busy restaurant kitchen, close up on hands and knife, steam in the background
-
Man jogging along a beach at sunrise, camera tracking beside him, waves in the background, sporty commercial look
All 5 models, ranked
Strength is the Plackett–Luce rating fitted to the judge's rankings, centred at 0; the bar is the 95% bootstrap interval. Overlapping bars mean a tie.
| # | Model | Strength | Ranks first | Prompt match | Looks | Credits per clipCredits |
|---|---|---|---|---|---|---|
| 1 | Wan 2.7 |
|
28% | 34.6 | 5.58 | 15 |
| 2 | Kling V3 |
|
26% | 37.8 | 5.26 | 20 |
| 3 | Seedance 2.5 |
|
25% | 37.0 | 5.41 | 35 |
| 4 | Veo 3.1 Fast |
|
15% | 36.3 | 5.23 | 12 |
| 5 | Veo 3.1 |
|
6% | 33.8 | 5.13 | 48 |
Narrow screens hide prompt match and looks, and show credits under each model name. Credits per clip are Picsart's own estimate on the benchmark settings.
Best for
Computed from the benchmark data, not written by hand.
Wins the most requests
2 of 4 first picks for Wan 2.7; 5 models won at least one.
Follows the prompt most closely
37.8 mean CLIP similarity: Kling V3, closest to the prompt.
Looks best to the aesthetic model
5.58 of 10 on the aesthetic predictor: Wan 2.7.
Most consistent of the top four
±0.69 log-strength: Seedance 2.5 has the narrowest error bar of the top four.
How stable the judge was
0.50 Kendall tau between two shuffled judge runs (1 identical, 0 unrelated).
What the judge saw
We judge three frames per clip: natural faces, hands and bodies, and no melting or extra limbs. Best first; the judge saw them shuffled and unlabelled.
-
A young woman dancing freely in a sunlit living room, loose white shirt, joyful, handheld camera, natural colors, social media video
- Wan 2.7 judge 1.5
- Kling V3 judge 2.0
- Seedance 2.5 judge 2.5
- Veo 3.1 Fast judge 4.0
- Veo 3.1 judge 5.0
-
chef chopping fresh vegetables quickly on a wooden board in a busy restaurant kitchen, close up on hands and knife, steam in the background
- Veo 3.1 judge 1.5
- Kling V3 judge 1.5
- Seedance 2.5 judge 3.5
- Wan 2.7 judge 3.5
- Veo 3.1 Fast judge 5.0
-
Man jogging along a beach at sunrise, camera tracking beside him, waves in the background, sporty commercial look
- Seedance 2.5 judge 1.5
- Veo 3.1 Fast judge 1.5
- Veo 3.1 judge 4.0
- Kling V3 judge 4.0
- Wan 2.7 judge 4.0
How this was measured
- Same requests. 4 people in motion prompts, written like real users write, sent to 5 video models from Picsart's AI Playground.
- Blind judge, twice. An AI judge ranked each request's clips without model names, from three frames per clip (at 25, 50 and 75%), then again reshuffled. It does not see motion or hear sound.
- Error bars. Plackett–Luce ratings bootstrapped over requests; CLIP and an aesthetic predictor on the same three frames shown as checks.
Questions people ask
- Which AI model is best for videos of people?
- Wan 2.7 (Alibaba): 28% chance of ranking first on 4 videos of people requests, ahead of Kling V3 at 26%, out of 5 video models tested the same way.
- Is Kling V3 as good as Wan 2.7 for videos of people?
- Close: their 95% error bars overlap (Wan 2.7 -0.66 to +3.08, Kling V3 -0.46 to +1.47 log-strength), so on 4 requests Wan 2.7 is only the more likely winner.
- How was the best model for videos of people measured?
- Same 4 requests to every model; a blind AI judge ranked the outputs twice, shuffled (agreement between runs: Kendall tau 0.50, where 1 is an identical order and 0 unrelated); Plackett-Luce ratings with 95% bootstrap error bars, CLIP and aesthetic scores as checks.