Skip to main content
Pika H3 Flex generates video clips with synchronized audio. A prompt is all it needs. You can also set the clip’s first frame, its last frame, or both from your own images. Or you can steer it with references: up to nine images for characters, objects, and settings, up to three audio clips for voices and sound, and up to three reference videos, which are experimental. To carry a scene forward, continue an earlier clip from its last frame, or from its last few seconds of picture and sound. You never choose a mode. The model picks how to build each clip from the inputs you send. Frame images and references cannot share a clip. Clips wait in a queue until they are ready, and then you play them. While one clip plays, you can queue more, change their order, or turn on autoplay so that ready clips play as they arrive. The Pika H3 Flex reference has four pages: this overview, the schema with every command and message, the prompt guide for writing scenes with references, and the tutorial, which walks through uploads, queues, and playback. Pika H3 Flex speaks the same wire protocol as every other Reactor model: you open a session with the Reactor class, send named commands, and receive messages. Its model name is reactor/pika-h3-flex. The @reactor-models/pika-h3-flex typed package adds named methods and React hooks on top of the same protocol.

At a glance

Key features

Video with sound

Describe the scene, the camera, the dialogue, and the soundscape in one prompt.

Frames and references

Start or end a clip on your own image, or steer it with reference images, audio, and video: characters, objects, settings, voices, and sound.

Clip control

Queue clips, choose the playback order, and continue a generated scene.

Quick start

Create a session token on your server, then connect with the base JavaScript SDK or the typed @reactor-models/pika-h3-flex package. Do not put your rk_ API key in browser code. See Authentication for how to create tokens.

How it works

  1. Connect to reactor/pika-h3-flex. The model describes the session with a state_update and a queue_update.
  2. If you want frame images or references, upload them first. Then call enqueue with a prompt. The model picks how to build the clip from what you send.
  3. Wait for clip_generated, then call play to stream the clip, or turn on autoplay instead.
  4. Queue more clips, change their order, or continue a clip that has finished generating.