Skip to main content
Back to timeline
arXivSource publication:

Streaming video editing drops inversion: AR-Edit caches source values once and edits each chunk as it arrives

Synopsis

AR-Edit edits video as frames arrive in an autoregressive diffusion model, without seeing the whole clip and without inversion. It exploits an observation about DMD-distilled models: self-attention queries and keys stay nearly fixed across denoising steps while values carry the changing content, so values cached from one clean forward pass are enough to anchor the source appearance. On the Reward Forcing backbone it starts from random noise at 6 network evaluations per chunk and scores 7.28 on source adherence, against 4.93 for the inversion baseline Inv, Inv, at 9 evaluations.

Source-provided article image: AR-Edit: Training-Free Autoregressive Video Editing without Inversion
arXiv

Interpretation

Streaming video editing can run without future frames, without inversion and without extra training, matching offline methods on source preservation and tonal fidelity. Earlier training-free editors needed the whole video and 20 to 50 network evaluations, while streaming approaches on the same backbones were limited to coarse noise-and-denoise editing with poor source preservation. On the Reward Forcing backbone, AR-Edit scores 7.28 on source adherence, against 4.93 for the inversion baseline Inv, Inv, and 3.81 for SDEdit; AR-Edit uses 6 network evaluations per chunk, the inversion baseline 9.

The method caches source values once and injects them selectively during denoising, while queries and keys keep the routing set by the target prompt. The inversion route first maps the source back into noise space with a multi-step integration and then re-denoises under the target prompt; with only 3 to 4 distilled steps that route breaks down. On Reward Forcing, a DMD-distilled student of Wan 2.1, queries and keys stay close to the clean anchor across the four training timesteps while values depart from it; the same asymmetry appears in the non-distilled bidirectional teacher.

Injection follows a Commit-then-Anchor schedule: full-frame injection in the first steps fixes the edit layout, then injection is confined outside the edited region and a text boost strengthens the target prompt. Full-frame injection throughout raises source preservation but suppresses the edit, while injection only outside the edit region keeps text adherence but recovers the source only partly. In four-step editing on Reward Forcing, the default configuration reaches 7.28 source adherence and 6.11 text adherence; removing Commit-then-Anchor drops source adherence to 5.68, and removing the text boost drops text adherence to 5.17.

Perspective

The demonstrated scope is streaming editing on four Wan 2.1-based models: three DMD-distilled autoregressive students (Reward Forcing, CausVid, Self Forcing) and the non-distilled bidirectional teacher; evaluation uses the authors' 120-clip Pexels benchmark across ten edit categories on a single H100.

The authors note that the query-key-value asymmetry is verified only on Wan 2.1-based models sharing one DiT design and has not been tested on structures such as MM-DiT that attend jointly over text and image tokens; the method inherits the source-versus-text trade-off common to training-free editors, and the injection step count, block threshold and text-boost weight have no universal setting, so more demanding edits may need per-example adjustment.