On October 18, 2024, Meta announced that it is open-sourcing Movie Gen Bench, the two benchmarks of the newly released media generation model Movie Gen. In early October, the tech giant released Movie Gen, a suite of AI models that can generate high-quality 1080p HD videos with different aspect ratios and synchronized audios.
To know more about the Movie Gen model, read here.
The Movie Gen models allow users to generate various types of media, including text-to-audio, text-to-video, and text-to-image. To accomplish this, it supports four models: Movie Gen Video, Movie Gen Audio, Personalized Movie Gen Video, and Movie Gen Edit models.
Meta introduced benchmark data in the Movie Gen model to facilitate its continuous evaluation and development. This includes two benchmarks: the Movie Gen Video benchmark and the Audio benchmark.
The Movie Gen Video benchmark is a comprehensive benchmark that facilitates the evaluation of text-to-video generation features. It consists of 1000+ prompts to understand the motion levels of human activity, nature, animals, or unusual objects.
Read More: Meta Introduces AI-Driven Assistant: Metamate
The Movie Gen Audio benchmark enables evaluation of video-to-audio and (text+video)-to-audio generation. It comprises 527 generated videos, relevant sound effects, and music prompts suitable for various ambiance.
By open-sourcing these benchmarks, Meta wants to ensure that the AI research community can leverage them to develop better audio and video generation models. It is in accordance with Meta’s commitment to open and collaborative AI development.
According to Meta’s X post, the tech company aims to promote fair and detailed evaluation in media generation research through the Movie Gen bench. For this, it has included diverse videos and audio generated from Movie Gen in the benchmark data.
To test Movie Gen’s performance, Meta recently partnered with Blumhouse Productions, a Hollywood production house known for horror movies like The Purge and Get Out.
