Pika Enact lets you direct AI actors and the camera live
The research preview generates video and sound in sub-second chunks, but Pika's own speed figures come from lab hardware and it is not for sale.

Key takeaways
- Pika published Pika Enact, a research preview built on Pika H3 Flex that generates video and sound in sub-second chunks so a creator can stage, direct and re-light a scene while it plays.
- Pika says Enact's live staging mode generates video 1.8 times faster than playback and responds in under 0.4 seconds on an H100, figures SLOP TV has not tested and that Pika does not give for every mode.
- Pika says Enact is unavailable for consumer use, with no price, waitlist or release date, and entry only through enterprise@pika.art.
- Fal's H3 Max Director, a comparable real-time video model on the same MiniMax H3 lineage, is callable now through fal's API at $0.08 per second of generated video, according to fal's own page.
Pika published a research preview it calls Pika Enact, a live-streaming environment the company says lets a creator stage, direct and re-light an AI scene while it plays. Pika posted it on its blog under Research and links it from its homepage. Both the post's header and the homepage card read November 18, 2026.
Pika says Enact runs on Pika H3 Flex, generating video and sound in sub-second chunks that carry scene context forward while folding in new direction. The goal, the post says, is directing "on set" instead of typing into a prompt box.
The first mode, Live Staging, lets a creator drag an actor or object into place with the mouse while the video keeps playing. Pika says the model generates the moved subject together with the rest of the scene, so pose, lighting and shadows adapt, and that the subject reacts to being moved before it settles. Pika reports it generates 1.8 times faster than playback on a single server, responding to input in under 0.4 seconds on an Nvidia H100, decoding included.
Live Actor Direction lets a creator type or speak a note and have the on-screen actors play it out. Pika says they answer in character, in their own voices, remember earlier notes, and that calling "Action" turns the rehearsal into one continuous take. Pika reports picture and sound generate 1.75 times faster than they play.
The third mode steers a virtual camera with the mouse while the scene generates, from an appearance image plus a rough 3D proxy such as an untextured Blender viewport. Pika reports that on a node with eight Nvidia B300 chips, at 640x480 and 24 frames per second, it renders the first frames in 0.51 seconds and later chunks at 1.50 times real time.
Those figures are Pika's own, measured on Pika's hardware. None has been run on an independent harness, and Pika gives no resolution figure for Live Staging or Live Actor Direction, and no latency figure for Live Actor Direction. Pika adds that its rehearsal demo was edited for length.
Enact sits on Pika H3 Flex, which Pika calls its post-trained version of MiniMax's H3 model and launched on October 29, 2026 with open weights. Nothing in the announcement says Enact itself is open.
The control has a cost. Pika CEO Demi Guo told The Neuron that the responsive, real-time model is slightly lower in image quality than Pika's non-real-time version.
Enact is not the only interactive video system. Fal's H3 Max Director, a comparable real-time model on the same MiniMax H3 lineage, is callable now from fal's API at $0.08 per second of generated video, with a 60-second minimum.
If Enact ships as a product, blocking, performance and camera work would sit under a mouse and a spoken note, and actors that keep their context would cut the retry loop. That is conditional: the speed figures come from data-center chips, Pika publishes no price, and the fastest camera number is at 640x480.
Enact is not for sale. Pika asks anyone who wants to try it or discuss an integration to email enterprise@pika.art.
Sources
- pika.art - the research preview post: all three Enact modes, their figures and conditions
- pika.art - H3 Flex specs: draft at 768p, finish up to 4K, up to 12 references
- pika.art - H3 Flex launch date (October 29, 2026) and the Enact homepage card dated November 18, 2026
- theneuron.ai - Demi Guo on the real-time quality tradeoff
- fal.ai - rival H3 Max Director: API access, $0.08 per second, 60-second minimum
- vp-land.com - the specs Enact leaves unstated; H3 Flex as post-trained MiniMax H3
- avenews.am - Pika's figures in table form; the post's conflicting dates