Face in a crowd
The prompt “A man standing still in focus in a busy train station, dozens of people walking behind him out of focus”
Tests whether background humans stay coherent. Usually they morph, merge or walk through each other.
- Models tested
- 15
- Top score
- 5.00 4 models tied
- Average
- 4.75
- Lowest
- 3.50 Seedance 2.5
Every model, same prompt
The clips on screen play together from the first frame, muted. Open one for sound, controls and our notes.
MiniMax H3
MiniMaxBackground pedestrians stay coherent, no merging or morphing.
MiniMax Hailuo 02
MiniMaxBackground crowd shows correct occlusion and depth blur, no merging or morphing.
Google Veo 3.1
GoogleBackground commuters stay separate and coherent, no merging or morphing.
OpenAI Sora 2
OpenAIForeground subject sharp; background crowd coherent, no glitching.
Google Gemini Omni
Google
Kling 3.0 Turbo
Kuaishou
Runway Gen-4.5
Runway
Vivix W1
VivixThe man stays still and sharp while an out-of-focus crowd crosses behind him in a train station. Background figures hold together with minor blending between passers-by. Checked frame by frame.
Kling 3.0
KuaishouSharp subject, coherent background crowd.
PixVerse v5.5
PixVerseBackground crowd faces render smudged/indistinct at distance -- typical, expected limitation of background-crowd rendering, not a specific defect.
Kling 4.0
KuaishouThe man stays still and sharp, but the crowd behind him is in focus rather than blurred. Checked frame by frame.
Wan 3.0
AlibabaThe man stays still and sharp, but the crowd behind him is only slightly softened rather than out of focus. Checked frame by frame.
FLUX 3
Black Forest LabsThe man stays still and sharp while a dense crowd passes coherently behind him, but the background is not out of focus as asked: the passers-by are nearly as sharp as the subject. Verified frame by frame.
Grok Video
xAISharp subject, coherent blurred crowd.
Seedance 2.5
ByteDanceFAILURE MODE: background pedestrians blur into translucent silhouettes and pass through one another instead of staying coherent. Foreground subject stays stable.
Ranked by our blind vision-model judge; models with the same score share a position. Read how we test.