EyeHeight benchmark reveals camera height regression in video world models
What happened
Sheng Lu proposed and released the EyeHeight benchmark, pointing out that six recent open-weight interactive video world models exhibit "camera height regression": after a 180° turn-and-walk, the models pull the camera height toward human eye level and cannot maintain the initial camera height. The benchmark includes six rendering heights from 0.15 m to 2.2 m and four calibrated real-camera heights, using scale-free multi-view height estimation. Starting from 0.15 m, only 0–20% of segments from the six models end within 1.3× the starting height, with the inferred viewpoint converging to 1.1 m to 1.5 m; ground-truth rendering error is only a few percentage points. The paper states that unguided sampling combined with height-balanced fine-tuning improves height retention but costs 5 MUSIQ points, a loss that is recovered within the same early steps.
AI-written from the coverage · updated 4 d ago
Coverage timeline
Follow the coverage to see the story from different sides.
- arXiv Game World Models & GenerationPickEyeHeight benchmark reveals camera-height regression in interactive video world models
The paper reports camera-height regression in six recent open-weight interactive video world models: starting from a height of 0.15 m and walking after a 180° turn, only 0-20% of clips end within 1.3× the starting height, suggesting the viewpoint converges to a human eye height of 1.1 m to 1.5 m, while ground-truth rendering error is only a few percentage points.
Heat trend for this story
- Current (comparable)
- 3
- Peak (comparable)
- 4Nov 18 11:00
- 24-hour change
- –
The trend only compares participants observed continuously and completely, so it may cover fewer than the current heat count. Hover over or tap the chart to see each hour's heat; with a keyboard, use the left and right arrow keys.