Field guide · verified 05 Sep 2026
Director turns a video request into a live session. Configure a world, receive video and audio over WebRTC, then steer upcoming segments with new prompt messages while the stream continues.
Hold on her face. Steam from the bag. She laughs and looks at camera.
Browser receives video and optional audio tracks.
Configure, ping, prompt, and stop.
The published schema reports requested chunk durations in this range.
fal marks its realtime client API alpha and subject to change.
01 / The family
Use the endpoint name as evidence. Do not infer Director from a generic H3 Max label—or call a MiniMax H3 project H3 Max.
One continuous session. It receives a detailed opening configuration, then compact prompt changes while the session keeps delivering media.
fal describes Max as a post-trained MiniMax H3 variant tuned for prompt adherence and aesthetics, co-optimized with its inference stack.
The speed-oriented clip path used by just-in-time projects such as UNREEL and Live Classroom, where the next segment renders during playback.
02 / Session lifecycle
The public Director schema is small. Product quality comes from what surrounds it: state, moderation, spend limits, recording, and fallback.
Create a realtime client through an authenticated server proxy. The browser receives media; the provider key never enters client code.
fal.realtime.open(wma("minimax/h3-max/director"))Send the world contract, prompt version, optional first-frame image, aspect ratio, resolution, seed, and memory.
{ type: "configure", prompt_version: 1, prompt }Send one new visible beat. Keep identity and world facts stable; increment prompt versions so the product can reconcile what applied.
{ type: "prompt", prompt_version: 2, prompt }Use configured, prompt-applied, chunk, deadline, error, and session metrics to drive truthful UI and fallback decisions.
chunk · prompt_applied · deadline_missedSend stop, close the peer connection, release recording resources, and settle the session server-side even if the tab disappears.
{ type: "stop" }03 / Price is a moving input
On 5 September 2026, fal's H3 Max text-to-video page showed temporary launch rates of $0.0125/s at 480p and $0.02/s at 768p, with stated post-promotion rates of $0.05/s and $0.08/s after 7 September. Director has its own session billing and product-level minimums may apply.
Time-sensitive source: fal H3 Max pricing panel, checked 05 Sep 2026. Verify the provider calculator before every budget or launch.
Check current fal pricing04 / In the wild
Native Director sessions and clip-buffer products belong to the same emerging category, but they should not share the same architecture label.
AI TV
fal's public experimental network streams themed AI channels and lets viewers shape what happens next.
H3 Max DirectorNative realtime
AI TV
A live public channel where viewers propose ideas, watch a generation queue, and vote on what they see.
H3 MaxJust-in-time generation
Interactive story
A fixed-format AI survival show where the audience directs four characters one shot at a time.
H3 Max TurboBuffered interactive
AI TV
A public viewing site that turns a premise into a queued H3 Max movie or episodic program and keeps completed titles in a browsable catalog.
H3 MaxJust-in-time generation
Interactive story
A public point-and-click story game where each choice creates the next H3 Max video segment and saves a persistent branch through the adventure.
H3 MaxJust-in-time generation
Interactive story
A public staging prototype where a visitor starts a story, chooses what happens next, and requests another 5–15 second H3 Max Turbo scene.
H3 Max TurboJust-in-time generation
05 / Questions
No. On fal, H3 Max is a clip-generation endpoint, while H3 Max Director opens a continuous WebRTC session that accepts new prompt messages while video and audio are streaming.
No. A product can be realtime because it continuously streams, generates faster than playback, or hides generation behind a buffer. Network, prompt application, generation, decoding, and playout still add delay.
No. fal's Director documentation says to keep the FAL_KEY on the server, route browser requests through an authenticated proxy, and protect that proxy from unauthorized spend.
Not yet. They are original, source-grounded structures and are labeled as RetivAI adapted and not independently generated by RetivAI.
fal Director API, fal H3 Max endpoint, and MiniMax H3 repository. Provider behavior and prices can change.