BLOG ILLUSTRATION / EXECUTION STORY

Change controls after each frame

AR-Drag compares two modes. Offline denoising receives a full control sequence. Causal generation keeps fixed history and receives changed guidance.

Change controls after each frameAR-Drag compares two modes. Offline denoising receives a full control sequence. Causal generation keeps fixed history and receives changed guidance. BIDIRECTIONAL · full window Full control sequence known before generation Joint denoising full frame window Output clip output after the window AUTOREGRESSIVE · next frame Current control Guidance can change. Causal denoising history + new guidance Next frame fixed output KV history
Author-redrawn AR-Drag timing comparison. The bidirectional mode uses controls for a full window that can change before output. The autoregressive mode keeps fixed history and receives guidance for subsequent frames. This diagram does not redraw the latency–FID panel. Its points, baselines, hardware and units stay in the original. Arrow spacing and playback time do not show runtime. Open full-size SVG