VidProm
收藏资源简介:
VidProm是一个用于生成视频的文本提示数据集,由人类编写的文本提示组成,旨在帮助模型更好地适应真实世界的人类输入。该数据集用于生成多个视频,并通过OmniScore评分系统进行质量评估,形成偏好对数据集。数据集的创建过程包括自动生成偏好对数据,并通过重新加权策略提高训练效率。该数据集主要应用于视频生成模型的用户偏好对齐,旨在提高生成视频的视觉质量和语义对齐。
VidProm is a text prompt dataset for video generation, composed of human-written text prompts, which aims to help models better adapt to real-world human inputs. This dataset is utilized to generate multiple videos, and conducts quality assessment via the OmniScore scoring system to form preference pair datasets. The dataset creation process includes automatically generating preference pair data, and adopting a reweighting strategy to improve training efficiency. This dataset is primarily applied to user preference alignment for video generation models, with the objective of enhancing the visual quality and semantic alignment of the generated videos.

- 1VideoDPO: Omni-Preference Alignment for Video Diffusion Generation香港科技大学, 中国人民大学, 约翰斯·霍普金斯大学 · 2024年



