A new arXiv preprint introduces EmoTrace-Bench, an evaluation framework for text-to-video (T2V) models that generate videos with targeted emotional expressions. The paper argues that current evaluation methods mostly rely on holistic video-text alignment, which can say whether a video generally matches a prompt but gives little insight into whether emotional expression itself is the problem.
EmoTrace-Bench addresses this by using controlled cues and intensity levels, allowing researchers to probe emotional expression in a more structured way. Instead of a single overall score, the benchmark is meant to diagnose where and how generation fails—whether the emotion is missing, too weak, too strong, or inconsistent with the prompt.
The work highlights a growing gap in T2V evaluation: as models are used more for emotionally specific content, generic alignment metrics become less sufficient. By focusing on controlled emotional cues, EmoTrace-Bench offers a more targeted tool for understanding model behavior and guiding future improvements.