fal has released the H3 Max Director mode for the MiniMax H3 Max video model. Instead of a set of individual clips, the model generates a single continuous video stream, maintaining characters, locations, and narrative continuity, while the prompt can be changed in real time, directing the action on the fly. The public fal.live demo is set up like “AI television,” where viewers vote on what happens next, and the API is available on fal.ai.
What Happened
Access to the mode is available through the minimax/h3-max/director API: the model is hosted in the fal.ai playground. Public sessions are currently limited to approximately two minutes, with longer ones enabled manually for approved cases. Billing is calculated in seconds of generated video: until September 14, a promotional price of $0.02 per second applies with a minimum billing unit of 60 seconds, so the first run costs $1.20. After the promotion, the base price will be $0.08 per second, and the same 60 seconds of stream will cost $4.80.
Context
fal is an inference platform where the MiniMax H3 Max API is sold. Until now, the standard format for video generation has been a short clip based on a fixed prompt: the model operates within the framework of a single clip, and maintaining the scene between generations is not set as a task for it. H3 Max Director changes the operating mode itself: context must be maintained over time, not within a single clip.
Why This Matters for the Industry
For the industry, this is a new class of workload for video diffusion/flow models: the model must maintain characters and scenes over time and respond to new prompts without breaks. This changes the product form — streaming, interactivity, “television” — and the economics of generative video: billing is tied to seconds of stream, not clips. H3 Max Director is the first public API for this mode with transparent pricing, and the promotional window makes prototypes of interactive video streams the cheapest in the category’s history.
Why This Matters for Users
Readers have direct access without development: on fal.live, you can watch an “infinite” AI broadcast and choose where the plot goes, and in the fal.ai playground — start a session independently. The mode is currently sold at a promotional price, making experiments almost free. A practical billing detail: a session shorter than a minute is billed as a minute, so short tests do not reduce the bill below the minimum threshold of 60 seconds.
What Is Still Unknown / Limitations
There are no articles, technical reports, or reproducible evaluations of the mode: quantitative metrics for character and location consistency, human evaluation protocols, and comparisons with the base MiniMax H3 Max are absent. Production characteristics — latency, throughput, stability of long sessions — are not disclosed, and there is no SLA. The limitation of public sessions to approximately two minutes may indicate real limitations such as drift and cost, which the vendor does not explicitly disclose, and the claim of scene continuity is currently confirmed only by demo clips.
Sources
- fal.ai — H3 Max Director model page (minimax/h3-max/director): price, limits, prompt examples
- fal.live — public demo “AI television directed by everyone”
Author
Look at AI, editorial team