A virtual set for directing scenes in real time with Pika Enact
Still experimental, Pika Enact combines subject placement, character direction, and interactive camera control in continuously generated video.
The word “Action” triggers a fresh take informed by the directions given during rehearsal. In Enact, the experimental system presented by Pika, characters can receive written or spoken instructions, respond in character, and perform the scene again. The final take starts from the opening frame: the system transforms the preparatory exchanges into a continuous performance with synchronized visuals and voices.
This continuously generating environment is built on Pika H3 Flex. It combines character direction, mouse-driven subject placement, and virtual camera control. The research overview specifies that Enact is unavailable for consumer use; trials and integration opportunities require contacting the team. The rehearsal demonstration was edited for length, with subtitles and the director’s webcam inset added afterward. Turning Directions Into Performance and Dialogue
A request to bring two characters closer together, change an emotion, or adjust the lighting becomes a set of structured instructions. A script language model reads the directions, opening frame, and session memory, then prepares the actions, dialogue, gaze, and atmosphere for the audiovisual model to perform. Generation continues throughout these exchanges.
Session memory retains feedback and rehearsal history to inform subsequent interactions. When no further instruction is queued, the system maintains subtle activity, such as breathing or blinking. For this directing mode, Pika reports audio and video generation at 1.75 times playback speed. That ratio describes generation throughput rather than the response delay for each spoken instruction. Using the Mouse to Shape the Scene and Framing
Live staging involves grabbing a character or object and moving it while the video continues. According to the team, the model regenerates the subject together with its surroundings, adapting posture, lighting, and shadows to the requested movement. In the low-latency configuration described, the image responds in under 0.4 seconds on H100 hardware, including decoding. This measurement concerns subject movement, using generated segments lasting 0.17 seconds.
Camera control uses a different setup. A reference image establishes the appearance of the characters and set, while a simplified 3D representation, such as an untextured Blender viewport, guides movement. Users can change the angle, move toward a subject, or orbit around it with the mouse. Audio and video are generated from these references and the scene’s history.
For this mode, the published test uses eight NVIDIA B300 GPUs at 640 × 480 and 24 frames per second. The first frames appear after 0.51 seconds; subsequent segments are generated at 1.5 times playback speed. A Stream Trained to Follow Gestures
All three modes generate short segments successively while maintaining a sliding memory of the scene. Distillation reduces the number of steps needed for each segment, and training teaches the model to continue from its own previously generated frames. An anchor image and recent history help maintain the appearance of characters and sets.
To teach mouse-driven movement, the team creates synthetic video pairs: an original scene and a version in which a 3D stand-in follows a trajectory, with cursor positions recorded. For camera control, a Blender production pipeline synchronizes trajectories, 3D references, images, and spatial audio. The document also describes a limitation of the opening anchor: during relatively inactive scenes, it can pull generation back toward the initial state. Periodic instructions are added to preserve changes in lighting or staging.