The generative livestream lets its audience take control

With fal.live, viewers can influence continuously generated videos in real time through H3 Max Director, an experimental version of H3 Max.

⚠️ EDIT, AUG. 31, 2026, 10:20 AM No livestreams are currently being broadcast on fal.live. The platform distinguishes this temporary pause from a technical outage: the generation engine remains operational while fal refines the system ahead of its next session. Broadcasts are expected to return soon, although no specific timeline has been provided. The experience therefore cannot be tested live at this time, despite fal stating that its infrastructure is functioning properly.

A show that transforms as its audience writes in the chat. That is the concept behind fal.live, an experimental platform dedicated to AI-generated video livestreams produced in real time. Viewers select a channel, submit instructions, and vote to determine what happens next.

Several formats have been prepared for the demonstration, including Chaos, Sitcom, Toon Lab, Choice, Tiny World, and Hot Mic. Each channel follows a different premise but relies on the same mechanics: the video continues to evolve as suggestions from the chat shape the next scenes. Users can also propose their own show concepts.

fal refers to the experience as H3 Max Live. Behind the interface, however, is another name: H3 Max Director. It is not exactly the same as the H3 Max endpoint already available for generating clips from text or images. Instead, it is an experimental version designed to extend generation over time.

According to fal Research member Batuhan Karaman, H3 Max Director operates autoregressively and is natively continuous. Each new sequence can build on previously generated material instead of starting over with an isolated clip. This approach is intended to produce smoother connections between scenes than conventional text-to-video or image-to-video workflows.

The context window can extend up to two minutes. This gives the system access to a longer video history to preserve a situation, character, or narrative direction. That duration does not mean the livestream has unlimited memory: fal has not yet explained what is retained, summarized, or gradually discarded beyond the two-minute window.

The experience relies on the speed of H3 Max, a version of MiniMax H3 post-trained by fal. According to the company, the standard model can process five seconds of video in under three seconds. Since processing takes less time than the resulting clip lasts, continuous broadcasting becomes possible with a shorter delay between an audience instruction and its appearance on screen. The official H3 Max presentation attributes this speed to the combined work on post-training and the inference engine.

The system does more than display instructions in the order they arrive. Suggestions can be put to a vote, while the highest-ranked questions may receive an answer directly within the show. Viewers therefore act less like individual directors and more like members of a collective writers’ room whose decisions are filtered through the system.

This format also introduces several challenges. Continuous generation must preserve characters, environments, chronology, and the audience’s intentions without allowing inconsistencies to accumulate too quickly. Moderation also becomes more difficult when public instructions can alter a broadcast within seconds.

At the time of verification, the channels on fal.live are temporarily paused. The page states that the generation engine remains operational, but that the system and its chat moderation safeguards are being refined. Broadcasts are expected to resume at a later stage, with no more precise timeline provided.

H3 Max Director is still described as an experiment. fal plans to improve its quality rapidly but has not yet published detailed documentation covering stream resolution, audio generation, resource consumption, broadcasting costs, or user limits.

No dedicated public endpoint for H3 Max Director is currently listed in the documentation. H3 Max remains accessible for text-to-video and image-to-video generation, while Director currently serves as the engine behind fal.live. The platform should therefore be viewed as a public demonstration of continuous video generation rather than a stable service ready to be integrated into other applications.