H3 Max DirectorDirect video as it generates.
H3 Max Director is the first natively continuous realtime frontier video model. It generates one unbroken video stream and holds its context over time, so you send new directions while it is running and the characters, setting, and story carry through. The clip below is one such stream: a single take with no cuts, in which the same character holds while the world around him changes completely. It is the model behind fal.live, and it is now callable from your own app. 75% off for the first two weeks: $0.02 per second of video until September 14, 2026.
One continuous H3 Max Director take, 30 seconds, no cuts. See the original post.
Three messages, and the stream never stops
Director is not a queue. A session is a WebRTC peer connection you hold, with the generated video and audio arriving as live tracks and your directions going back on a data channel. Everything below happens inside one open connection.
What continuous generation gets you
One stream, not a chain of clips
Director generates without stopping, so there is no seam where one generation ends and the next begins.
- Each segment continues from the last through 39 frames of carried context.
- A small re-rendered overlap is dropped before the video reaches you, so the join is invisible.
- Segments run 10 seconds by default, and can be set anywhere from 5 to 15.
Change the story without restarting it
Send a prompt mid-session and the action turns to follow it, with nothing regenerated and nothing restarted.
- A direction is applied to the next segment while the current one is still playing.
- The change lands as a cut in the story rather than a break in the stream.
- The server confirms the moment your direction reaches the generator, so you can show it landing.
Characters and settings hold across every turn
Director expands each new direction against the prompts that came before it, which is what makes a cast stay itself. It is the same continuity H3 Max holds inside a single clip, extended across a session that never ends.
- Prior segment prompts are kept as context, 12 by default and up to 50.
- Faces, wardrobe, geography, and the running plot survive a direction that changes the scene entirely.
- A higher memory holds a longer story; a lower one lets the stream turn more sharply on each direction.
Sound generated with the picture
Audio is generated in the same pass as the video and arrives on its own live track, already in sync.
- 32 kHz audio against 24 fps picture, both streaming from the same session.
- Dialogue, foley, score, and room tone are part of the generation rather than added to it afterwards.
- Describe the sound in the same prompt as the shot and it arrives with the scene.
What people are building
A stream you can steer, and a model that renders faster than its own playback, turn out to be the two things interactive video needed. These are builds on Director and the wider H3 Max family, each one noting which it runs on.
The story is the input
Offer a choice, send the option taken as the next direction, and the picture follows it. The branch lands as the next beat of the same scene instead of a load screen, and the world, cast and sound already built up survive the turn.
A stream the audience changes
On an ordinary livestream the chat has no effect on the picture. Here it is an input to it: a host can take questions and work the answers into the broadcast while it runs, and viewers can push prompts or images that change what is on screen. With the right IP agreements in place, the same pattern covers branded characters doing cameos, greetings or product explainers.
Direct out loud
A direction is a text message on the session's data channel, so anything that can produce text can drive it, speech to text included: say what happens next and the stream turns at the next segment.
fal.live is one Director session per channel
fal.live is fal's experimental infinite AI livestream, and every channel on it is a single H3 Max Director session running continuously. Steering is collective: the stream offers the room a choice, viewers vote, and the winning option is handed to the Director while the current scene is still playing, so the story turns without the video ever stopping. One session serves a channel's whole audience, so viewer count adds no GPU load at all.
Fifty worlds to direct
Every channel below is a real Director preset, with the world prompt its session opens on. Copy any of them and it runs verbatim, or open one in the playground and steer it somewhere the preset never mentioned.
- Comedy · Live actionDirect a Live Sitcom
Drop a fixed ensemble into audience-created misunderstandings.
News Satire · BroadcastAbsurd Breaking NewsTurn audience prompts into an escalating fictional breaking-news broadcast.
Adventure · GameIslands With Impossible RulesHelp a shipwreck survivor navigate islands with impossible rules.
Competition · Reality TVAround-the-World RaceSend recurring racers through an ever-changing world-spanning course.
Anime · AnimeMagical Girl MishapsA heroic magical-girl broadcast whose transformations keep going wrong.
News Satire · AnimeAnime Breaking NewsCover the day's fictional headlines as a continuous anime broadcast.
Fantasy · AnimeInfinite Anime QuestDirect a spiraling anime quest where every detail can become a world.
Comedy · Reality TVImpossible Cooking OrdersMake a painfully serious chef obey ridiculous cooking instructions.
Mystery · CinemaRain Memory MysteryInvestigate memories carried by the rain in a neon port city.
Comedy · DocumentaryDog Park Power StruggleA prestige mockumentary about alliances and scandals at a dog park.
- Sci-fi · AnimeBroke Space Bounty Hunters
A broke bounty-hunting crew crosses the galaxy in a failing luxury liner.
Reality TV · AnimationFruit Dating ShowProduce romantic chaos among single fruits in a tropical villa.
75% off for the first two weeks
Director bills by the second of video it generates, not by the request, because a session generates for as long as you hold it. There is no subscription and no minimum commitment beyond the per-session floor.
- Right now
- $0.02 / second
- From September 14, 2026
- $0.08 / second
- Minimum per session
- 60 seconds
- A full 2-minute session
- $2.40
75% off, until September 14, 2026
List price, per second of video generated
A session shorter than that still bills $1.20
$9.60 once the launch discount ends
Public sessions run up to 2 minutes. Longer Director sessions are being enabled gradually for approved use cases: tell us what you are building and our team will get back to you. Full platform pricing is on the pricing page.
H3 Max endpoints
Director is the realtime one: it holds a session and streams. The rest of the H3 Max family answers a request with a finished clip, which is what you want when you need a file rather than a stream. Browse the wider catalogue on the video models page. Information updated as of September 7, 2026.
Direct continuous, realtime video streams with live prompts while preserving characters, settings, and story continuity.
fal's H3 Max is a post-trained variant of MiniMax H3, tuned for stronger prompt adherence and better aesthetics while co-optimized with our custom inference stack for higher throughput with no compromises on output quality
fal's H3 Max is a post-trained variant of MiniMax H3, tuned for stronger prompt adherence and better aesthetics while co-optimized with our custom inference stack for higher throughput with no compromises on output quality
fal's H3 Max is a post-trained variant of MiniMax H3, tuned for stronger prompt adherence and better aesthetics while co-optimized with our custom inference stack for higher throughput with no compromises on output quality
fal's H3 Max Turbo is a post-trained variant of MiniMax H3, tuned for stronger prompt adherence and better aesthetics while co-optimized with our custom inference stack for higher throughput with no compromises on output quality
fal's H3 Max Turbo is a post-trained variant of MiniMax H3, tuned for stronger prompt adherence and better aesthetics while co-optimized with our custom inference stack for higher throughput with no compromises on output quality
Call H3 Max Director from your own app
A session is opened over WebRTC rather than submitted to a queue, so the client that plays the stream is the client that holds it. Everything after the handshake is JSON on a data channel. The transport underneath is WMA, fal's realtime interface for world models, which any app on fal Serverless can build on.
import { createFalClient } from "@fal-ai/client";
import { wma } from "@fal-ai/client/realtime";
const fal = createFalClient({ proxyUrl: "/api/fal/proxy" });
// `configure` opens the world. It is the first message on the channel and
// carries everything fixed for the session; only the prompt changes after it.
const world = {
type: "configure",
protocol_version: 1,
prompt:
"A continuous original live-action stream following a group of friends as they explore a new city.",
prompt_version: 1,
resolution: "768p",
aspect_ratio: "16:9",
memory: 12,
};
const session = fal.realtime.open(wma("minimax/h3-max/director"), {
// The client receives video and audio and sends none: the stream is the
// output, and your directions go back over the data channel.
receive: ["video", "audio"],
onState: (state) => {
// The data channel does not exist until the peer is live, so the world
// is sent here rather than straight after open().
if (state === "live") session.send(world);
},
onMedia: (stream) => {
const video = document.querySelector("video");
if (video) video.srcObject = stream;
},
onData: (raw) => {
const message = JSON.parse(raw);
console.log(message.type);
},
onError: (error) => console.error(error),
});
Common questions about H3 Max Director
What makes H3 Max Director different from other video models?
A request/response video model, H3 Max included, answers with a finished clip. Director holds a session instead. It generates one continuous stream and keeps its context across time, so the video does not end and restart at a clip boundary: each segment continues from the last through 39 frames of carried context. That is what lets you change the story while it is playing rather than between generations.
How much does an H3 Max Director session cost?
Sessions bill $0.02 per second of video generated, a 75% launch discount that ends September 14, 2026. After that the list price is $0.08 per second. Every session bills a minimum of 60 seconds of runtime, so a session shorter than a minute still costs a minute.
How long can a Director session run?
Public sessions run up to 2 minutes. Longer sessions are being enabled gradually for approved use cases: if you need to hold one for longer, get in touch through the form at the bottom of this page and tell us what you are building.
What resolutions and aspect ratios does Director support?
480p or 768p, at 16:9, 9:16, or 1:1, generated at 24 fps with 32 kHz audio. All of them are fixed when the session is configured and hold for its whole run, because changing them would mean starting a new stream.
How does Director keep characters consistent across directions?
It retains prior segment prompts as context, 12 by default and up to 50, and expands each new direction against them before generating. A higher memory holds a longer running story; a lower one lets the stream turn more sharply on each new direction. It is set once, when the session is configured.
Can I build a game or an interactive experience on Director?
Yes, and it is the thing Director is best at. Present the viewer with a choice, send the option they take as the next direction, and the stream turns to follow it while the current segment is still playing, so the branch arrives as the next beat of the same scene rather than a load. Because prior segments stay in context, the world, cast and sound you have built up survive the turn. Directions are plain text on the session's data channel, so anything that can produce text can drive it, speech to text included: this prototype directs by voice, paired with a reasoning model, though it runs a modified Director build rather than the public endpoint. The limit to design around today is session length: public sessions run up to 2 minutes, with longer ones enabled per use case.
What is fal.live?
fal.live is fal's experimental infinite AI livestream, built on H3 Max Director. Each channel is one Director session running continuously, and viewers steer it: prompts submitted on the page are moderated, voted on, and the winner is handed to the Director while the current scene is still playing. It is the reference implementation for what this endpoint does, and the model underneath it is the one on this page.
Can I call Director from Python?
A Director session is a WebRTC peer connection rather than an HTTP request, so it is opened from the client that will play the stream, which in practice means the browser and the JavaScript client. A Python backend still has a job: host the fal proxy route the browser client points at, so your fal key stays on your server instead of shipping to the page.
Longer sessions with H3 Max Director
Public sessions run up to 2 minutes. We are enabling longer Director sessions gradually for approved use cases, so tell us what you are building and our team will get back to you.