Introduction

Clip-in, clip-out video APIs just got a rival with a pulse. fal’s H3 Max Director opens a WebRTC session, streams picture and sound as live tracks, and lets you send fresh directions while the scene is still playing—no queue hop, no restart, no awkward stitch.

It is the engine under fal.live, fal’s experimental AI “TV,” and it is now callable as minimax/h3-max/director from your own app. For creators who already love MiniMax H3, this is the moment the model stops being a file factory and starts feeling like a camera you can talk to.

Why Director is not “another T2V endpoint”

Classic video models answer a request with a finished clip. Director holds a session.

  • You open a WebRTC peer (fal’s WMA realtime layer), not fal.run / the queue.
  • Video and audio arrive as live tracks; your prompts ride back on a data channel.
  • Each segment carries 39 frames of context into the next, so joins stay invisible.
  • Prior prompts stay in memory (12 by default, up to 50), so faces, wardrobe, and plot survive wild turns.

Send a new direction mid-stream and the story cuts—the picture does not. That is the whole trick.

How a session feels in practice

Three beats, one unbroken connection:

  1. Open the peer with @fal-ai/client alpha (fal.realtime.open + wma("minimax/h3-max/director")).
  2. Configure the world once: prompt (up to 50k characters), 480p or 768p, aspect (16:9 / 9:16 / 1:1), memory, audio bitrate.
  3. Direct with later prompt messages while the current segment still plays.

Segments default to about 10 seconds (tunable 5–15). Audio is 48 kHz stereo generated in the same pass—dialogue, foley, score, room tone riding with the frame at 24 fps.

Keep FAL_KEY on a server proxy. The browser holds the peer; Python’s job is protecting the key, not opening WebRTC itself.

Pro Tip

Treat Director like a runtime, not a renderer. Build a choice UI (vote, click, whisper-to-text), ship the winning option as the next prompt_version, and let the cast you already established carry the branch. Public sessions top out around 15 minutes—design for clean reopen rather than infinite single takes.

fal.live and the builder wave

fal.live is the reference channel: one Director session per stream, audience prompts moderated and voted, winner handed to the model while the current scene still rolls. Viewer count does not multiply GPUs—one session serves the room.

Around it, builders are already shipping interactive horror maps, synthetic “press play and it writes itself” film shelves, and point-and-click mysteries. The pattern is the same: generation time under playback time turns spectators into co-directors.

Pricing window (act before Sep 14)

Director bills by generated video second, with a 60-second minimum per session:

WindowRate

Now through September 14, 2026

$0.02 / sec (75% off)

After Sep 14

$0.08 / sec list

A full 15-minute public session is about $18 on the launch price, $72 afterward. That is channel economics, not hobby-clip math—but for interactive demos, live shopping, adaptive ads, and narrative games, the unit economics suddenly make sense.

Where it sits beside other H3 flavors

Director is the realtime sibling. fal still ships request/response H3 Max and H3 Max Turbo endpoints when you need a finished file. Think of them as studio renders; Director is the live set.

If you already follow Reactor-style realtime H3 stacks elsewhere, keep them for their niches—Director’s pitch is the open WebRTC session + prompt memory + fal.live proof in one fal-shaped package.

Conclusion

Video gen is growing a nervous system. H3 Max Director gives you a steerable world with synced sound, character hold across turns, and a public stage on fal.live—plus a short discount window before the meter jumps.

Open a session, configure the world, and whisper the next beat while the lights are still up.

—Aurelia ♡