❗️❗️❗️**LEADERBOARD INTRODUCTION:** ❗️❗️❗️ This is A2-Bench leaderboard which is used to evaluate the performance of elements-to-video (E2V) generation models. We provide an evaluation set containing 50 paired multiple elements (character, object, and background). You can check [evaluation set introduction]() for more details. Each evaluation case includes:

• Human subject (characters): Includes both male and female subjects, covering generated by Flux, celebrities and ordinary people, additionally, we provide several generated human images
• Non-human subject: Various objects including different types of animals, guitars, racing cars, balls, etc.
• Background image: Diverse environmental settings including ordinary indoor and outdoor scenes and famous background suck as The Great Wall and Yellow Wind Ridge (from Black Myth: Wukong)
• Prompt: "A realistic scene where [human] interacts with [object] in [environment], following physical laws and spatial logic".

We provide a set of evaluation metric of elements-to-video models and a leaderboard to show the performance of different models. Evaluation metric include: - Elements Consistency: Measures character id consistency using arcface human recognition model, and measures object and background consistency using CLIP model. - Video Quality: Measures video quality on image quality, dynamic degree, aesthetic quality and motion smoothness. - T2V Metrics: Measures text-video consistency using CLIP You can check [Metric Introduction](https://skyworkai.github.io/skyreels-a2.github.io/static/images/bench.png) for more details. The leaderboard ranks the models based on the comprehensive score, which is the weighted average of all the metrics. We give T2V metrics and object consistency metrics higher weights. You can click the model name to visit the project page, At meantime, you can upload your model result **zip file** to the leaderboard by clicking the "Evaluation" button. the zip file should include the video result named as "0.mp4", "1.mp4", "2.mp4", etc.