@reactor-models/sana-streaming SDK. By the end you’ll know how to edit your webcam feed in real
time, steer the edit mid-stream, snap clips, and surface model errors.
Installation and setup
Get the example running before reading further. Every section below points back at code in the example repo. You will need:- Node.js 18+.
- pnpm (the example’s lockfile is pnpm’s;
npmoryarnwork too). - A Reactor API key (starts with
rk_). - Familiarity with the Next.js App Router.
1
Clone the example
The example lives alongside our other reference apps in
reactor-team/reactor-cookbook under
examples/api-models-examples/.2
Add your API key
Your
rk_… key must never reach the browser; the example reads it server-side and mints a
short-lived JWT for the client (the standard broker pattern); for now, drop
the key into .env.local:3
Install dependencies and start the dev server
http://localhost:3000, click Connect, allow camera access, pick a preset prompt (or
type your own edit), and press Start live.How SANA-Streaming works
Building with SANA-Streaming is different from Reactor’s other models. Helios, LingBot, and LongLive-2.0 generate video from a prompt; SANA-Streaming edits the video you bring. You open a long-lived connection, publish your webcam and an edit instruction, and the model streams back transformed frames in 24-frame chunks, paced to your camera’s own rate. Re-prompt at any time and the new edit lands at the next chunk boundary, with no re-render and no break in the stream. Opening the connection isn’t instant. Reactor provisions a GPU for your session, so the client moves through four states before media starts flowing:waiting state is when the GPU is being assigned, which takes a few seconds. Once the status
reaches ready, commands take effect and the
session lifecycle begins in its idle
state. StatusBadge.tsx surfaces every connection state with a label and a Connect / Disconnect
toggle. See Sessions for the full breakdown.
Two properties of the API are worth internalizing before you read on:
- Commands are asynchronous; messages are the source of truth. Calling
setPromptdoesn’t mean the edit changed yet; the model confirms with aprompt_acceptedmessage and astatesnapshot carrying the newcurrent_prompt. - Errors arrive out-of-band. A broken precondition like
resumewith nothing paused surfaces later as acommand_errormessage (useSanaStreamingCommandError), not a thrown exception.
@reactor-models/sana-streaming SDK, which wraps the base SDK with
the model name and tracks baked in. SanaStreamingApp.tsx mounts
<SanaStreamingProvider getJwt={fetchToken}>; components read status and call one typed method per
command off useSanaStreaming() (setPrompt({ prompt }) rather than
sendCommand("set_prompt", { prompt })), and subscribe to messages with per-message hooks like
useSanaStreamingState. The wire shapes behind every method and message are in the
schema.
Clip recording is the one exception. It is model-agnostic and not re-exported by the typed
package, so
SnapClip.tsx uses the base @reactor-team/js-sdk
directly (useReactor, <ClipPlayer>). Drop it into any model example unchanged.The model is the source of truth
The browser sends commands and renders the state the model reports back; it never tracks generation state on its own. That discipline lives in one reducer inapp/lib/state.ts, which projects the
model’s typed state messages into a small SanaState. Because it is fed only by the
useSanaStreamingState hook, the reducer never has to filter by message type, and it reads the
snapshot’s flat, typed fields directly:
app/lib/state.ts
Workspace shell in SanaStreamingApp.tsx subscribes with three typed message hooks:
useSanaStreamingState feeds the reducer, and useSanaStreamingCommandError and
useSanaStreamingGenerationReset handle the two messages that need side effects, imperatively:
app/SanaStreamingApp.tsx
pause with nothing generating, resume while not paused);
the shell turns each command_error into a banner that dismisses itself after six seconds.
Every control in the app gates off the reduced SanaState: the Start button on state.started, the
pause/resume and reset controls on state.started and state.paused. Informational messages
(prompt_accepted, chunk_complete) are not state inputs; whatever they report also arrives in the
next state snapshot, which is the
canonical payload.
The same discipline shapes the commands going out. A command only takes effect once the model echoes
it back in a state snapshot, so the start path confirms nothing itself: it calls start() and
lets the reducer report when generation is running.
Editing your webcam
Send your webcam to SANA-Streaming by publishing your camera to the model’scamera input track,
then call start(). Edited frames come back on the main_video track about a second later.
LiveInput.tsx owns the camera acquisition rather than reaching for a declarative webcam component:
app/components/LiveInput.tsx
publish and unpublish from useSanaStreaming() in an effect keyed on the track
and the connection status: it publishes once the session is ready, re-publishes after a reconnect,
and unpublishes on unmount. The typed SDK also ships a declarative <SanaStreamingCameraView> that
acquires and publishes the webcam for you, but it gives no hook to set contentHint, so LiveInput
owns the track and publishes it by hand.
The Start live button calls start() and is disabled until status === "ready", the publish
has resolved, and no generation is running. A run has no length limit: it keeps transforming the
camera until the user resets it or the session ends.
The stage
Stage.tsx renders the model output with the typed <SanaStreamingMainVideoView>:
app/components/Stage.tsx
<ReactorView> with track="main_video" pre-bound, and manages the <video> element,
srcObject binding, and browser autoplay quirks for you. Apply your styling to the container around
it, not to the video element it renders.
A status row along the bottom reads running / paused, currentChunk, and currentPrompt
straight off the reduced state.
After a reset the model emits nothing new, so the view would freeze on the last transformed frame;
the shell blacks the stage out until state.running flips back to true.
Steering the prompt mid-stream
Prompts are editing instructions, not scene descriptions: “apply a Van Gogh oil painting style,” not “a Van Gogh painting of a room.”Prompt.tsx is one textarea, one Apply button, and a row of preset
chips, and every path funnels into the same call:
app/components/Prompt.tsx
setPrompt works before start and at any point mid-stream; the model applies it at the next chunk
boundary. A prompt is optional: start without one and the model streams a near-reconstruction until
you set a prompt to steer it. The textarea’s placeholder (“Describe the edit. Changes apply live,
about one chunk later.”) spells out the latency. The active-prompt readout under the button renders
state.currentPrompt, so it reflects what the model is using rather than what was last typed.
The preset chips come from app/lib/examples.ts. They are deliberately short style tags (“Van Gogh
oil painting, swirling brushstrokes, vivid colors”) that show how a whole-frame restyle reads against
a live feed, leaning on the model’s default to carry everything else through. For surgical edits, a
garment swap or a removal where you must spell out what stays fixed, write the fuller instructions
the prompt guide lays out, with its anatomy and a
recipe per edit type.
Playback, seed, and reset
The Input panel (ModeInput.tsx) is phase-aware, driven by the model’s started flag. Before a run
it shows the webcam input and the seed field; once started flips true it keeps the input slot
mounted (so the camera keeps publishing) and swaps the setup controls for Playback.tsx: pause,
resume, and reset. Each is a typed method off
useSanaStreaming(), gated on the reduced state:
app/components/Playback.tsx
SeedField.tsx, a setup-phase control that calls setSeed({ seed }) on blur. The
model reads the seed at start, so the same source, prompt, and seed reproduce a run; the field is
keyed to the model-reported seed, so a reset or an external setSeed refreshes it with the model’s
value.
reset does the most work: it aborts the run, clears the model’s prompt and progress, and emits
generation_complete and generation_reset. The shell’s handler (from
The model is the source of truth) mirrors that on the client.
It blacks out the stage and clears the prompt draft, so the UI matches the model.
What’s intentionally left out
The demo covers the connect + edit + steer + capture loop. Clip capture is a shared base-SDK feature, so Recordings covers it, including continuous recording, programmatic capture, and retention. A few other patterns are out of scope, and each is a small addition:
For the full design rationale and the patterns to follow when extending the app, including the
typed-SDK surface and the manual camera publish, read
skill/SKILL.md in
the example repo.