Cut.ProCut.Pro

Pika Enact: directing AI video live, like on a film set

On Wednesday (Nov 18), Pika showed Enact, a research preview where you drag actors, give direction notes by voice and move the camera while the scene is being generated. Here's how it works, what's still missing and what you can use today.

Pika Enact: directing AI video live, like on a film set

Pika Enact: directing AI video live, like on a film set

On Wednesday (November 18), Pika showed Enact, a research preview that generates video and sound in real time and lets you direct the scene as it happens: drag the actor, give direction notes by voice, move the camera. It isn't open to the public, and Pika itself says it isn't available for consumer use.

Still, it's worth understanding what it shows. Today's workflow is to write a prompt, wait minutes and hope for the best. Enact swaps that for rehearsing, correcting and calling "action," and that's the direction video generators are heading.

What does Pika Enact do?

It generates the scene in short chunks and accepts commands in the middle of it, in three modes described in Pika's announcement:

  • Live staging. You drag an actor or object with the mouse while the video plays, and the model adjusts pose, light and shadow to the new spot.
  • Live actor direction. You give notes by typing or speaking. Each actor is an AI agent with its own personality, answers in character with its own voice and remembers what you said in earlier takes of the same session. When you call "Action," they do a new, continuous take.
  • Live camera. You move a virtual camera with the mouse (pan, zoom, orbit) while the scene is generated. A reference image holds the cast and the set in place.

The Rundown, which reported on the preview, described Enact as built to work like a live film set.

How Pika got to real time

The trick is not generating the whole clip at once. According to Pika, Enact produces video and audio in chunks of about 0.7 second (0.17 second in the low-latency staging mode), and each chunk is generated from what has already come out, without waiting for the rest of the scene. In staging, the company also cut the number of steps the model needs to draw each chunk from twenty to just a few.

The numbers Pika published:

Mode What Pika measured
Live staging response under 0.4 second on one H100 GPU, generating 1.8x faster than real time on a server
Actor direction image and sound 1.75x faster than playback
Live camera 640x480 at 24 frames per second on a server with 8 NVIDIA B300 GPUs, first frames in 0.51 second

Two practical takeaways from this table. First, 640x480 is low resolution and horizontal: it doesn't work directly for Reels or TikTok. Second, a server with eight B300 GPUs to run the live camera is lab-level cost. That's why Enact sits on a research page and not in a subscription plan.

What about H3 Flex, which already exists?

Enact is built on Pika H3 Flex, the model Pika released on October 29. That one you can actually use: it generates clips of up to 15 seconds with audio, in 480p or 768p, and accepts up to 12 reference files across image, video and sound. It launched first in the Pika API, with the creation app planned for November 2, and the open weights were published for anyone running their own server.

In the technical report, Pika lists a starting price of US$ 0.01 per second generated from text or image and US$ 0.03 per second in reference mode. The same report is honest about the limits: in physics, multi-scene storytelling, lighting and materials, and in high-action tasks, other models were preferred in Pika's own study. The Rundown also notes that the actor demo video was edited to fit, so there's no way to know how it all works together in a long session.

What this changes for video creators

Today, anyone generating scenes with AI spends more time redoing than creating. The prompt gets the character's position wrong, you tweak a word, wait again and pay again. Making a character enter from the left instead of the right can take several generations.

Enact goes straight at that: instead of rewriting the prompt, you drag the character to where they should be and keep going. If the idea catches on, three things change:

  1. The cost of a mistake drops. Fixing something mid-take is cheaper than generating the whole clip again.
  2. Direction becomes speech. Saying "slower, look at the camera" is more natural than finding the right word for a prompt. For anyone rehearsing a sketch or a short story, that brings AI closer to how you direct people.
  3. Generating becomes a format. Just as walking through a playable Genie 4 world can produce exploration videos, directing AI actors live could become a livestream or behind-the-scenes format on its own.

The catch: none of this counts until Pika opens access, and there's no date or confirmation for Brazil. There's also no word on vertical video, which is what matters for short-form.

What to do now

Don't wait for Enact to fix your workflow. Three steps that work today:

  • Generate cheap before generating pretty. Enact's principle (fail fast at low resolution) is something you can already apply: test the scene at the lowest resolution, which usually costs less per second, and only redo the take you got right at a higher resolution. In the Cut.Pro Generator you pick the model, generate paying with credits and take the clip straight to the editor.
  • Use a reference instead of rewriting the prompt. A reference image holds character and set from one take to the next, which is what the anchor image does in Enact. You can see this in the face photo animation test and in the AI persona guide.
  • Don't bet your channel on one tool. The end of Sora 2 showed that a dazzling preview doesn't guarantee a product. For per-second pricing on a model that's already live, see Gemini Omni 1.2 Flash.

Enact isn't in the Generator and isn't open to the public, so this post comes without a real test. When Pika opens it up, I'll test it and update this post.

Share

Keep reading

More insights and tutorials to help you grow as a content creator.