Back home

Pika Enact

sitemap+web 0.93
Visit source

Intro

Pika Enact is the new name Pika introduced as a research preview for an experimental environment that lets a creator direct a generating video in something closer to a live film set. This routine first noticed the official page in the sitemap diff for already-tracked sites. The blog says Enact is built on the already-filed video model Pika H3 Flex: it generates picture and sound in sub-second chunks, keeps scene context, and folds new directing input into the clip that is still playing.

Three modes are described. Live Staging lets you drag actors or objects while the video keeps playing, with placement, lighting, and shadows adapting to the surroundings instead of stopping for a fresh render. Live Actor Direction lets you talk to the “actors” about dialogue, emotion, or movement, then call “Action” to start a continuous take that remembers those notes. Live Camera Control is mouse-directed camera movement around a scene that is already playing.

Official copy frames this as a research preview, not a feature already on for every consumer; public coverage has said interested users should contact Pika’s enterprise team. This entry records the new name Enact, not H3 Flex itself.

Advertisement

Explain

Pika presents Enact as a step toward AI filmmaking that looks more like directing on a set than typing into a prompt box. Technically it streams low-latency generation on top of H3 Flex, taking mouse trajectories, spoken notes, and camera moves as continuous input rather than one prompt and one finished clip. The blog says that in its low-latency setting, the picture can respond in under 0.4 seconds on an H100, decoding included, and that generation can run about 1.8× faster than playback on a single server so a drag does not have to stall the video.

Live Staging is described as needing tightly aligned before-and-after video pairs with frame-level cursor trajectories, which real footage rarely provides. The team built a physics-driven synthetic data engine: a 3D stand-in is moved along a drag path on an otherwise untouched clip, photorealistic targets are rendered, and mouse positions are logged automatically. That yields training pairs in which only the dragged subject changes, covering lifts, carries, and inserting a new subject from a photo into a live scene.

Actor direction and camera control are the other two inputs in the same real-time loop, including giving notes and turning rehearsal into a continuous take. Access, waitlists, and whether this lands in the product should be checked on the official blog and Pika’s current statements. Enact is not just another text-to-video model name; it is a live directing environment stacked on H3 Flex.

Usage

Read Research Preview: Pika Enact for the three modes and demos, then follow whatever contact path the page gives for enterprise or research access. Do not assume a consumer generate button already includes it. To understand the underlying model, see the already-filed Pika H3 Flex post, but judge Enact on whether live dragging, spoken direction, and camera moves actually keep the scene continuous at sub-second feedback.

Practical questions include whether the surroundings close up after you move a person, whether actors remember earlier notes, and whether a camera move breaks an already-stable identity. Latency and hardware figures are from the official research page; independent checks depend on access being granted.

If you only want Pika to generate a clip from a prompt, that is existing Pika model capability, not this new research-preview name.