A detailed survey of training datasets for image-to-video generation models has been published, covering major systems such as Stable Video Diffusion (SVD), Wan, CogVideoX, HunyuanVideo from Tencent, Runway Gen-3 Alpha, Kling from Kuaishou, and Open-Sora. The report compares dataset sizes, sources, preprocessing pipelines, and licensing considerations. For example, SVD uses a large-scale video dataset with diverse motion patterns, while HunyuanVideo leverages Tencent's internal data. This survey is a practical resource for AI teams looking to understand the data landscape for video generation, identify gaps, and make informed decisions about dataset curation. It highlights the importance of data diversity and quality in achieving high-fidelity video outputs.
This post surveys training datasets used by major image-to-video models including Stable Video Diffusion, CogVideoX, HunyuanVideo, and others. It provides a valuable reference for researchers and engineers building or fine-tuning such models, comparing data sources, sizes, and preprocessing approaches.