A Video is Worth 256 Bases：零样本视频编辑的时空 EM 反演

一句话结论

这篇工作把时空 EM inversion 推到零样本视频编辑里，补强了 video-editing 的 inversion / reconstruction 路线。

问题定义

它要解决的是零样本视频编辑中如何更稳地反演和操控视频表示的问题。对当前知识库来说，它补的是 video-editing 里与图像编辑 inversion 路线相呼应的一支。

方法概述

该工作通过 spatial-temporal expectation-maximization inversion 来支持 zero-shot video editing，让视频内容更好进入可编辑状态。

关键发现

它说明 inversion 并不是图像编辑独有问题，视频编辑也在发展自己的时空反演路线。
它和 NULL-Text Inversion、Effective Real Image Editing 形成跨图像/视频的机制对照。
它有助于以后比较“训练式编辑”和“零样本反演式编辑”的边界。

局限或疑问

零样本反演方法通常仍然受效率和稳定性约束。
它更偏反演机制，不覆盖更开放的高层语义编辑。
它是机制层锚点，但不等于完整视频编辑系统。

原始链接

论文原文（CVF OpenAccess）：https://openaccess.thecvf.com/content/CVPR2024/html/Li_A_Video_is_Worth_256_Bases_Spatial-Temporal_Expectation-Maximization_Inversion_for_CVPR_2024_paper.html
官方 PDF（CVF）：https://openaccess.thecvf.com/content/CVPR2024/papers/Li_A_Video_is_Worth_256_Bases_Spatial-Temporal_Expectation-Maximization_Inversion_for_CVPR_2024_paper.pdf
arXiv：https://arxiv.org/abs/2312.05856
arXiv PDF：https://arxiv.org/pdf/2312.05856
代码仓库：https://github.com/STEM-Inv/STEM-Inv

备注

这篇工作在库里的意义，是把 video-editing 的时空 inversion 路线补成明确节点。

Metadata

{ "id": "2026-04-14-video-worth-256-bases", "type": "source", "title": "A Video is Worth 256 Bases：零样本视频编辑的时空 EM 反演", "status": "reviewed", "created": "2026-04-14", "updated": "2026-04-15", "venue": "CVPR 2024", "ingested_at": "2026-04-14", "tags": [ "near-cvpr-2025", "video-editing", "diffusion", "primary-source" ], "note_status": "reviewed", "source_type": "paper", "authors": [ "Li, Maomao", "Li, Yu", "Yang, Tianyu", "Liu, Yunfei", "Yue, Dongxu", "Lin, Zhihui", "Xu, Dong" ], "published_at": "2023/12/10", "canonical_links": [ "https://openaccess.thecvf.com/content/CVPR2024/html/Li_A_Video_is_Worth_256_Bases_Spatial-Temporal_Expectation-Maximization_Inversion_for_CVPR_2024_paper.html", "https://openaccess.thecvf.com/content/CVPR2024/papers/Li_A_Video_is_Worth_256_Bases_Spatial-Temporal_Expectation-Maximization_Inversion_for_CVPR_2024_paper.pdf", "https://arxiv.org/abs/2312.05856", "https://arxiv.org/pdf/2312.05856", "https://github.com/STEM-Inv/STEM-Inv" ], "raw_entry": "raw/ingest/2026-04-14-video-worth-256-bases/", "topics": [ "topics/video-editing", "topics/diffusion-models" ], "entities": [ "entities/video-editing-understanding" ], "claims": [], "questions": [ "questions/question-do-benchmarks-track-real-video-editing-understanding" ] }

A Video is Worth 256 Bases零样本视频编辑的时空 EM 反演