The first useful question is not “which model is best?”

1. An open session: H3 Max Director

Provider says: fal’s H3 Max Director keeps a WebRTC session open, sends audio and video tracks to the client, and accepts text directions while the session runs. This is the route that most closely matches “the audience changes what is happening now.”

The conditions matter. The current provider page lists 480p or 768p output, a 60-second minimum charge, $0.02 per generated second until 14 September 2026 and $0.08 afterward, and public sessions up to 15 minutes. A 15-minute cap does not kill the idea; it makes reconnect and story handoff part of the product.

Unknown: this pool has not measured time to first frame, time from a new direction to visible change, actual billing, or what survives a reconnect. A provider learning article dated two days earlier also describes a two-minute default, so recheck endpoint capabilities before treating any session limit as fixed.

2. Buffered clips: FastH3

Project report: Reactor’s FastH3 implementation separates prompt generation from playout. It builds short clips into a ready queue, then a client decides when to play one. This is valuable because it shows the system most launch posts skip: queue, buffer, playout, failure behaviour, and delivery.

The project reports a warm 14.4-second build for a 14.375-second clip on four B200 GPUs. That is a capacity signal under a declared high-end profile, not an affordable general recommendation. The same notes require CUDA 13 and roughly 148 GB of weights, and warn that new shapes can bring compilation delay. A stream that lives on this route needs ready footage ahead of its viewers.

3. Local short clips: MiniMax H3

Provider says: the open MiniMax H3 base covers 4–15 second clips at 24 fps, with separate FL2VA and Ref2VA variants. That is a real local path for controlled clips. It is not an endless-generation claim.

Community report: a recent RTX 4060 Laptop report describes an 8 GB H3 FL2V configuration at 480×864 and 158 frames. The author reports a 10–11+ minute wait at a larger 544×960 configuration. That is encouraging for an offline experiment and decisive against calling the same workflow a live feed.

Version is part of the setup. An open ComfyUI issue reports a full-resolution H3 run becoming much slower on later versions in one RTX 5070 Ti environment. That one report does not settle the cause, but it makes “H3 on a 16 GB card” an incomplete answer.

The practical answer

Use Director when the experience needs an active, guided session and you can accept a provider bill and a session-limit design. Use a FastH3-style queue when you control the schedule and can keep finished clips ahead of playout. Use local H3 when the task is a short local clip and you are ready to capture exact conditions. No source inspected here establishes a low-cost, self-hosted system that can run continuously without a buffer or a high-end reference profile.

Next evidence that would change the decision

1. A measured Director session: connection time, direction delay, bill, session end, and reopen behaviour.
2. A reproducible local baseline at stated H3 version, ComfyUI version, GPU, resolution, frames, and steps.
3. A smaller FastH3 profile with measured ready-buffer performance and rental cost.