Efficient-Large-Model/SANA-Streaming-example-training-dataset
收藏资源简介:
该数据集包含1000个对齐的反向视频编辑对,用于公开的SANA-Streaming双向视频到视频(V2V)训练配方。具体内容包括500个背景更改、167个局部添加、167个局部移除和166个局部更改。每个样本由源视频(.source.mp4)、目标视频(.target.mp4)和元数据文件(.json)组成,数据集以ZIP分片形式组织,每个分片包含100对。数据集用于非商业研究用途,支持SANA-Streaming模型的微调训练,但请注意这是一个示例子集,并非完整训练混合数据。
This dataset contains 1000 aligned paired reverse video edit pairs, developed for the publicly released SANA-Streaming bidirectional video-to-video (V2V) training recipe. It specifically includes 500 background modification cases, 167 local additions, 167 local removals, and 166 local modifications. Each sample is composed of a source video (.source.mp4), a target video (.target.mp4), and a metadata file (.json). The dataset is structured as ZIP shards, with each shard holding 100 pairs. This dataset is intended for non-commercial research purposes only, and supports the fine-tuning of the SANA-Streaming model. Please note that this is an example subset rather than the complete training mixture data.




