license: cc-by-4.0
configs:
- config_name: prompts
data_files:
- split: latest
path: prompts.jsonl
Dataset Card for Amazon Nova Reel Evaluation
This document contains some additional context on the settings and methodology for how we evaluated the Nova Reel model. Unless stated otherwise, we use greedy decoding with the maximum generation length set at 1600 tokens.
Responses from the Nova models are released under a CC-BY-NC license. The public benchmarks used to generate responses can be accessed via the hyperlinks provided below, subject to the applicable license terms for each benchmark
Nova Reel Human Evaluation Prompts
We curated a diverse set of prompts designed to capture various aspects of video generation. The prompts are distributed across 6 broad categories: human and activities, animals, natural scenery and landscapes, indoor scenes, objects interactions, and creative scenes and activities. This broad categorization ensures that the evaluation covers a wide range of real-world scenarios. We structured the prompt set to cover various motion-related aspects, which is critical for assessing motion-text alignment in the generated videos. For example, we included prompts with a variety of camera motions to evaluate how well the models follow instructions related to camera movement. Additionally, we incorporated dynamic attributes, in which the subject or background undergoes state or shape changes over time, which allows us to evaluate the model’s ability to generate evolving entities. Finally, we added prompts that require motion binding, where specific compositions of movements and actions are requested, enabling us to assess how well models can generate complex, coordinated motions. The curated prompt set consists of approximately 700 prompts, all from various open source benchmarks.