FateZero
FateZero is a zero-shot framework for text-driven video editing using pretrained diffusion models without requiring additional training or user-specific masks. Recognized as an ICCV 2023 Oral presentation, this tool enables consistent editing of real-world videos by fusing intermediate attention maps captured during DDIM inversion to preserve structural and motion information. Unlike standard approaches, FateZero directly integrates these attention maps during the editing process rather than generating them anew. It further reduces semantic leakage from the source video by blending self-attentions with masks derived from cross-attention features of the source prompt. The system also reformulates the self-attention mechanism in the denoising UNet by introducing spatial-temporal attention to ensure frame consistency. FateZero supports diverse use cases including text-based style transfer, local attribute modification, and shape-aware editing. It outperforms previous methods in temporal consistency and editing c