To Their Credit

Video

Pika previews Enact, an AI video tool built for live direction

The research system lets creators move performers, give spoken notes and steer virtual cameras while a scene is generated.

Ledger: AdoptedAI is taken up

Pika has unveiled Enact, a research-only AI video environment intended to let creators direct a scene while it is still being generated. Announced Nov. 18, the system generates video and sound in sub-second chunks, retains scene context and incorporates new instructions without requiring a finished clip before changes can be made.

The company describes the project as resembling a live film set. Enact is built on Pika H3 Flex, which Pika announced Oct. 29, and is not available to consumers. The company has not disclosed Enact pricing and directs prospective users and integration partners to its enterprise team.

Its Live Staging feature allows a creator to drag an actor or object to a new position with a mouse as the scene plays. Pika says the system regenerates the subject and its surroundings, adjusting such details as pose, lighting and shadows rather than simply placing a cutout over a background. The company reports picture response of less than 0.4 seconds on an Nvidia H100 processor, including decoding.

Live Actor Direction accepts typed or spoken notes about movement, dialogue, emotion and setting. A script language model reads those instructions alongside the opening frame and session memory, then translates them into structured beats covering action, dialogue, emotion, gaze and lighting. Pika says actors can retain direction and feedback within a session.

The workflow also includes a rehearsal mode. Enact logs each line verbatim during rehearsal; when a creator calls “Action,” it generates a new, continuous performance from the first frame rather than editing interruptions out of the rehearsal footage.

A separate Live Camera Control mode permits panning, dollying and orbiting around subjects with a mouse while audio and video are generated anew. Pika reports synchronized audio-video chunks of about 0.71 seconds and steady-state generation of 0.473 seconds per 0.708-second chunk, or 1.5 times real time. Its camera demonstration used 640-by-480 resolution at 24 frames per second on eight B300 graphics processors.

To make rapid staging possible, Pika says it streams mouse input and decoding in small pieces and uses a distilled, few-step model. It also built a synthetic-data engine that moves 3D stand-ins along drag paths, producing before-and-after video pairs and records of mouse positions and button states. Camera training requires frame-level synchronization of 3D camera motion, video and spatial acoustics.

The system has safeguards intended to contain drift during a session: an opening anchor frame can pull idle scenes toward their initial state, while periodic hold beats preserve changes to lighting or staging. But the project’s practical value will depend on whether it can sustain responsive camera control and reliable retention of direction through a full session, according to The Rundown.

Magica noted that consumer access, production reliability and pricing remain unresolved. For now, Enact remains a research preview rather than a publicly usable video-generation product.

More from Video

All Video news
  1. Seoul AI Content Festival pairs generated media with dealmakingLedger: Adopted
  2. Zheng Yuanjie turns his fairy tales into free AI animation on DouyinLedger: Adopted
  3. Google opens Genie 4 interactive worlds to AI Ultra subscribersLedger: Adopted