Benchmark
StepDiff
An evaluation benchmark focused on fine-grained perception and description of differences between videos, examining a model's ability to identify and describe how two videos differ with respect to specific concepts.
Timeline (last 90 days)
Oct 20
VICE was evaluated on two benchmarks, ConViS-Bench and StepDiff, and outperformed models that process two videos simultaneously.
Unverified50%
Oct 20
The StepDiff benchmark focuses on fine-grained perception and description of differences between videos.
Unverified50%