Smooth or Jarring? Evaluating Video Transitions with TransiSense and VT-Bench

With the rapid advancement of state-of-the-art AI tools and diffusion-based models for scene interpolation, qualitative analysis has become increasingly time-consuming and tedious. Here, we address this challenge by proposing a novel metric, TransiSense, which offers an improvement over current benchmarks in the field. Existing evaluation methods primarily focus on isolated aspects of video quality, failing to comprehensively assess the overall transition between an initial and a final frame. Our proposed metric overcomes these limitations by providing a more holistic evaluation, serving as a valuable tool for researchers developing video generation models. Additionally, we introduce a new dataset, VT-Bench, designed to facilitate the testing and benchmarking of future models and metrics. The dataset can be accessed directly at https://huggingface.co/datasets/Abhirup04/VT-Bench .