H3 Max Director: steer the next scene.

The official fal playground provides sample scenes and a sign-in flow for starting a session. Running the model uses your fal account and its applicable credits or billing. This guide provides free planning tools; the playground is a separate service.

Start with a recognizable place and two fictional characters. Let the first moment establish them, then ask for a small change: a visitor arrives, an object moves, or a sound interrupts the conversation. Rewriting every detail at once makes it harder to judge how well the session remembers the scene.

Before starting, decide how many attempts you can afford. A session that lasts only ten seconds still incurs the minimum described below. Use the Director session calculator to compare repeated short trials with fewer longer ones.

On September 5, 2026, fal lists $0.02 per generated second as a promotion expiring September 14; the standard rate is $0.08/sec. Each session is billed for at least 60 seconds. Access to sessions longer than two minutes requires an application.

These examples multiply the published rate by the greater of generated seconds and 60. Two separate 30-second sessions therefore cost twice as much as one 60-second session. The totals exclude delivery and other operating costs. Confirm the applicable rate in your account before starting.

This is an architecture choice. A lower clip rate alone does not tell you which interaction will feel better. For a controlled comparison, use the same cast and setting, record time until playable output, and count how many results you actually keep. The model guide and stream budget cover the clip route.

The endpoint is minimax/h3-max/director. fal's current alpha SDK uses fal.realtime.open for WebRTC video and audio. The documented control messages include configure, prompt and stop. Prompt context defaults to 12 prior segments, with a range of 1–50; it is not a permanent story database.

The official API reference includes a Next.js server-proxy example. Keep the provider key on that server and authenticate the proxy before deployment. Pin and test the SDK version because this interface is still alpha.

For your own application, establish a single session owner. Disable duplicate starts while a connection is opening, give the user a clear end-session control, and record the final usage independently of what the browser displayed. A disconnected browser should not silently create another paid session.

Use two layers when writing directions. The first defines what should stay recognizable: names, appearance, room layout, tone and relationships. The second describes the next visible action. Our Director prompt builder creates both as separate, editable pieces of text.

For example, the world can be an orbital diner with a copper robot cook and a tired astronaut. The next action can be the cook noticing that a pancake has started floating. Keeping the cast, counter and lighting familiar gives that small event room to land.

Store important plot facts in your own application if the story must survive a session restart. A record of who found a key, which door is locked, or what the audience chose is useful even when the next generated shot changes an incidental detail.

A WebRTC session is a media connection, not a complete public broadcasting operation. An OBS scene or a server-side media bridge can form part of your delivery design, but test capture, audio synchronization, output encoding and reconnect behavior together. Destination streaming requirements still apply.

If you need to review every frame before publication, a queued clip workflow is easier to inspect than immediate playout. If you choose a continuous session, define how quickly the operator can stop delivery and what viewers see during an interruption.

Do not budget a day-long Director show as one unrestricted session. First obtain the required access and establish how session endings, restarts and billing work for your account. The builder guide connects these choices to a shared channel architecture.

1. Start one session and confirm that repeated clicks do not start another.
2. Change one scene detail and note when the direction appears on screen.
3. Mute, unmute and end the session using keyboard controls as well as a pointer.
4. Disconnect the network and check what the viewer and server each report.
5. Compare provider usage with the budget reservation, including any new-session minimum.

Keep the first test small enough that you can watch the entire result. Once one complete interaction works reliably, add a single audience decision and test again. Several people sharing a scene can make a different product from several independent private sessions; the cost model should describe the one you actually operate.