FateZero: Fusing Attentions for Zero-shot Text-based Video Editing

Chenyang QiXiaodong CunYong ZhangChenyang LeiXintao WangYing ShanQifeng Chen

Chenyang QiXiaodong CunYong Zhang ...+3 Qifeng Chen

Mar 2023

摘要

The diffusion-based generative models have achieved remarkable success intext-based image generation. However, since it contains enormous randomness ingeneration progress, it is still challenging to apply such models forreal-world visual content editing, especially in videos. In this paper, wepropose FateZero, a zero-shot text-based editing method on real-world videoswithout per-prompt training or use-specific mask. To edit videos consistently,we propose several techniques based on the pre-trained models. Firstly, incontrast to the straightforward DDIM inversion technique, our approach capturesintermediate attention maps during inversion, which effectively retain bothstructural and motion information. These maps are directly fused in the editingprocess rather than generated during denoising. To further minimize semanticleakage of the source video, we then fuse self-attentions with a blending maskobtained by cross-attention features from the source prompt. Furthermore, wehave implemented a reform of the self-attention mechanism in denoising UNet byintroducing spatial-temporal attention to ensure frame consistency. Yetsuccinct, our method is the first one to show the ability of zero-shottext-driven video style and local attribute editing from the trainedtext-to-image model. We also have a better zero-shot shape-aware editingability based on the text-to-video model. Extensive experiments demonstrate oursuperior temporal consistency and editing capability than previous works.

展开全部

图表提取