The system uses a three-stage distillation pipeline: foundation tuning gives a bidirectional diffusion transformer editing ability, causal adaptation converts it to chunk-wise streaming attention, and DMD distillation compresses generation to four diffusion steps. An autoregressive-oriented mask cache reuses tokens from unchanged regions to reduce redundant computation.
LiveEdit is useful for interactive video effects, creative tools, virtual production, and low-latency visual editing research. The project reports 12.66 FPS streaming inference and provides code, comparison galleries, and examples such as changing fur, lighting, clothing, sky, and other scene attributes while maintaining temporal continuity.
