Adobe previews real-time steering for AI-generated video
By Steven Van ·
MotionStream, an Adobe Research preview (not yet in Firefly), lets creators drag objects and adjust cameras mid-generation.
Adobe Research has previewed MotionStream, an experimental technique that lets creators steer AI-generated video while it's being made, rather than waiting on a full render and starting over to change it. Starting from a text prompt, users can click and drag objects to control their movement, adjust the camera position, and choose which elements stay static, with results updating in real time.
The system generates video in pieces rather than rendering a full clip up front, using an autoregressive model where each new segment depends only on what came before it. That lets the tool stream a video to the user as it's created and take feedback mid-generation instead of requiring a fresh prompt for each change. Adobe says the underlying model captures physics and natural motion, so an edited object (their example: an elephant) moves and reacts realistically without manual rigging or keyframing.
This is a research preview from Adobe Research, not a shipped feature in Adobe Firefly itself.
