TL;DR of the paper, generated automatically.

This arXiv paper presents SemantiScene-Edit, which edits video from a text instruction while keeping the source footage.

• Architecture: A multimodal encoder reads the source frames and the instruction, and passes them to a pretrained video diffusion transformer. A second path feeds source-video latents alongside the noisy target for spatial and temporal detail.
• Training: Flow matching, a three-stage image and video curriculum, and over 1.2 million instruction-conditioned editing records covering addition, removal, replacement and attribute changes.
• Evaluation: A 100-case benchmark scoring instruction execution, source preservation, temporal consistency and visual quality. The authors report it follows instructions while retaining source content.
• Input: At inference, it needs only a source video and a natural-language instruction.