fal Worlds is a new discovery and access surface for interactive world models. Its October 15 announcement includes H3 Max Director, the continuing-video member of fal's MiniMax H3 Max family. The Worlds launch does not mean Director itself was released that day: the model already had a public page and earlier demonstrations.
Use fal Worlds to find the experience and the Director endpoint contract for implementation details. This guide checked those sources on October 16, 2026. It describes the documented protocol and illustrative application behavior, not a paid session or project latency test. H3Vid has not integrated Director's live transport.
Find the right execution path
The model identity is minimax/h3-max/director. Its transport is WMA over WebRTC: a browser holds a peer connection, receives video and audio tracks, and sends control messages. The endpoint documentation explicitly says to avoid fal.run, fal.subscribe, and the Queue API for opening this session.
That distinction matters even if your application already generates ordinary H3 Max clips. A queue request eventually yields a downloadable file. A Director session delivers ongoing media and takes new directions while it runs. Converting those live tracks into a persistent MP4, recovering interrupted playback, and deciding when a session ends are separate application responsibilities.
The fal inference overview describes the broader WMA infrastructure. Different worlds on the same platform can expose different model contracts. A common landing surface does not make Director, LingBot, or another world model interchangeable.
Configure, acknowledge, then direct
The current guide uses the experimental fal JavaScript client and server proxy. Keep the provider key on the server and expose an authenticated proxy for the browser. Do not transplant an unauthenticated example route into a public application without its access checks.
After opening the connection, send one configure message with the world premise, protocol_version: 1, and prompt_version: 1. Choose the session's aspect ratio, resolution, and memory here. A transport state of live means the channel exists; it does not mean the model has accepted configuration.
Wait for configured carrying the matching version before enabling direction controls. Then submit a prompt message using version 2, increasing the version for later submissions. Keep one sequence across the session. A direction changes future generated material; it cannot rewrite frames already delivered to the viewer.
An original premise could be an archivist walking through a rainy harbor toward an observatory. A later instruction might ask the archivist to stop under a lamp and examine a map while preserving the coat and weather. Send that change, observe its acknowledgement, and wait to see it before introducing a contradictory action. This is a planning example, not a tested recipe.
Acceptance is different from visible playback
The protocol distinguishes prompt_pending, prompt_applied, and prompt_rejected. Match each event's version to the input that produced it. Applied means accepted for generation; the result may still be preparing or waiting to appear in the stream.
Newer updates can replace pending preparation, and an older input may never receive a final event. Mark that outcome unknown rather than silently calling it rejected. A useful interface can show the last submitted direction, the latest accepted direction, and what the viewer is currently seeing as separate states.
Likewise, a diagnostic error is not automatically a terminal session failure. The endpoint lists session-ending errors and input-specific failures separately. Handle a known input failure against its version; display unfamiliar codes without guessing which submission failed. This keeps recovery behavior tied to actual provider evidence.
Frames, soundtrack, and a planned sequence
The current schema includes optional opening and ending images, startup audio, and scripted beats. These are more specific than a generic image-reference upload. The opening image is described as an exact first frame; startup audio is FL2VA target audio whose source PCM plays, rather than a Ref2VA style reference.
The script path uses whole-second offsets on the generated video's clock and has combination rules. For example, top-level script cannot be combined with top-level ending image or audio; those belong inside the script. Follow the current schema before mixing them. Do not infer that an ordinary H3 Ref2VA prompt accepts the same controls.
Resolution and aspect ratio are session settings. The checked schema includes 480p, 768p, and 1080p, whereas the older Director marketing page describes only the first two. Use the current endpoint contract for accepted fields, and treat the older page's examples and historical promotion as dated material.
Budget the session, including its floor
The endpoint copy still displays a promotional $0.048 per generated second and states that the discount ends October 15, followed by $0.08. On this October 16 check, budget from the stated post-promotion rate and confirm it in fal before opening a paid session. A 60-second minimum makes the standard-rate floor $4.80; the stated 1080p multiplier doubles that arithmetic to $9.60.
These are calculations from published terms, not observed charges. Do not reuse the marketing page's expired $0.02 launch offer. The older page also describes a fifteen-minute public ceiling and prepaid-balance limits; verify the current allowance for your account rather than treating that ceiling as included credit.
Close the peer when finished. A closed connection, terminal failure, or stream_exhausted ends that session attempt. Start a fresh version sequence for a new session and avoid automatically replaying uncertain directions. For the distinction between sustained throughput and audience response time, see the H3 realtime AITuber explanation; for finished clip generation, use the separate H3 Max page.