Infinite Streaming Video Editing with InfinityEdit

Summary & Methodology Analysis

InfinityEdit addresses the problem of applying unbounded edit instructions to continuous video streams by leveraging a frozen Helios-Distilled backbone. By keeping the backbone static, the system preserves established priors for stable, long-duration video generation. The methodology centers on an edit-ignition adapter, an auxiliary module inserted after each transformer layer, which integrates history cross-attention, temporal causal self-attention, and edit cross-attention mechanisms. This design allows the model to inject instructional guidance without requiring a full re-train of the primary generation engine. The training process employs a two-phase curriculum that balances broad coverage with granular detail refinement using a flow-matching objective. Inference is handled via an ignite-then-continue strategy: upon receiving an instruction, the adapter is activated solely for the first chunk to initiate the edit. Following this, the model resets its anchor frame and updates the history window, enabling the frozen backbone to maintain continuity throughout the subsequent video stream. This architecture excels in directional camera control, achieving a Camera Motion score of 0.7654, which significantly outperforms the SANA-Streaming baseline by 0.22. Furthermore, the system displays exceptional consistency, maintaining edit faithfulness near 3.8 across the entire sequence with a standard deviation of 0.023, avoiding the high variance observed in competing methods. Despite its performance, InfinityEdit has notable limitations. The model currently accepts only natural-language instructions, lacking support for visual references such as target images or video clips. Additionally, the transition between specific instruction types can result in abrupt visual shifts, presenting a challenge for seamless, long-term generation.