Adobe: real‑time control for AI video
Adobe Research demonstrated MotionStream, a system that lets creators click, drag and redirect AI‑generated video in real time instead of waiting for repeated regenerations. The approach replaces a slow trial-and-error loop with interactive steerability over motion and composition, potentially changing how teams direct AI video output. (redsharknews.com)
Artificial intelligence video tools usually make a clip all at once, then force users to regenerate it if motion or framing is wrong. Adobe Research said on April 10 that its MotionStream system instead lets creators steer a video as it is being generated. (research.adobe.com) Adobe showed users dragging objects with a cursor, painting movement paths, and changing camera angles with sliders while the scene kept updating. The company described MotionStream as an experimental technology from Adobe Research rather than a shipping Creative Cloud product. (research.adobe.com) In the underlying paper, the researchers said current motion-controlled video systems often take minutes per video and cannot respond interactively because they generate fixed-length sequences offline. MotionStream, they wrote, reaches sub-second latency and up to 29 frames per second on a single graphics processor. (arxiv.org) The basic problem is speed. Most video diffusion models work like a batch printer, producing many frames together, while MotionStream is built more like a live feed that can keep extending a clip as new control inputs arrive. (arxiv.org) To do that, the team started with a text-to-video model that could follow motion guidance but not run live, then distilled it into a causal student model that predicts frames in sequence. The paper says that shift enabled streaming inference instead of waiting for a full regeneration pass. (arxiv.org) Adobe’s project page says the system can generate arbitrarily long videos from a single image and support track control, motion transfer, drag operations, and three-dimensional camera control. The demos shown on the page are presented as raw screen captures without post-processing. (joonghyuk.com) The work also picked up academic attention beyond Adobe’s own announcement. OpenReview shows the paper was accepted as an oral presentation at the International Conference on Learning Representations 2026. (openreview.net) The author list spans Adobe Research, Seoul National University, Carnegie Mellon University, and Morpheus AI. A public GitHub repository says code release is still under internal company review for open-sourcing. (github.com) That leaves MotionStream, for now, as a research preview of a different way to use generative video: less like submitting prompts to a render queue, and more like directing motion while the shot is still unfolding. (research.adobe.com)
