Researchers have released Code2Video-Compass, a benchmark designed for general-purpose code-based video generation. The benchmark spans three generation paradigms: text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V).
The work argues that code-based video generation offers advantages over direct generation, including fine-grained control, targeted editing, and component reuse. However, this approach requires multimodal large language models (MLLMs) to interpret visual and temporal information, which the benchmark is intended to evaluate.
The abstract is truncated, so details on dataset size, evaluation metrics, and baseline results are not available from the source. The announcement does not compare against other benchmarks.