One Editor, Many Edits: A Unified Training-Free Framework for Diverse Video Editing

EditVid is a training-free framework that turns a frozen image-editing multimodal diffusion transformer into a unified video editor for style transfer, attribute and part modification, object insertion, and reference-guided subject replacement. Built on FLUX.2-Klein-9B, with cross-backbone testing on FLUX.1-Kontext, it separates temporal consistency into three mechanisms: sparse causal memory reuses key–value states from the immediately preceding frame for local motion coherence; correspondence-based post-attention token injection transfers identity and appearance from an anchor frame over long ranges using confidence and cycle-consistency filtering; and soft latent blending preserves instruction-irrelevant regions without rigid masks. Using four denoising steps, EditVid reaches 78.16 FiVE-Acc, substantially above the strongest evaluated training-free baseline, FlowDirector, at 58.95, while remaining competitive on IVEBench and leading most criteria in a curated 50-video VLM evaluation. In a blind study with seven competing methods, it receives 51.8% overall preference. The best temporal-scope ablation uses only the immediately preceding frame, reaching 77.50 FiVE-Acc and 91.63 MF-S; retaining all previous frames instead degrades performance. On an NVIDIA H200, the system achieves 3.15 seconds per frame, demonstrating a practical accuracy–runtime trade-off without video-specific training or test-time optimization.