How to hold a scene across fifteen-second clips

Three fields decide everything

The public API takes one free-text prompt up to 50,000 characters. Underneath, the H3 family was trained on a labelled three-field description, and fal’s own prompt rewriter emits that shape in expanded_prompt. Writing it yourself removes a layer of guessing.

A no-people bumper

Clips without faces or hands fail least often, which is why an early channel should lean on idents, bumpers, weather, and wildlife.

integrated_multimodal_description: [Shot 1] Live-action, macro, shot on 1990s broadcast video.
A chrome kinetic sculpture rotates once on a black seamless backdrop.
Push In, small amplitude, slow speed. No people, no on-screen text.
No speckles, no floating artifacts, no extra limbs, no distorted faces.
overall_soundscape: One low mechanical servo hum, one metallic clack, faint room air.
non_diegetic_music: Four notes on an electric piano at a slow tempo.

• Description: 300 to 450 words per fifteen-second clip, only what is visible and audible, in playback order. Style contract first, then palette, texture, and light.
• Shots: two to four per fifteen seconds, none shorter than three seconds. One camera move per shot, named from a fixed vocabulary.
• Dialogue: stable speaker ids, one or two very short lines per clip, each inside one shot. Dialogue is the least reliable channel.
• Soundscape: two to four sentences naming concrete sources and materials. Never describe music here.
• Music: one or two sentences of instruments, tempo, dynamics. Write N/A to disable, then never mention music again.

From three words to four hundred

A separate language model converts viewer input into the full format in one pass. Slop TV’s published system prompt is the clearest design: impose a hook, build, and payoff; add one turn the viewer did not ask for; moderate by steering rather than refusing.

• No worked examples in the system prompt: a small model parrots concrete examples.
• When nobody is waiting, feed RANDOM plus recently aired titles and forbid reusing those subjects.
• Never let the queue empty. A recycling LIFO queue capped at 100 messages gives the GPUs work forever.

Two official mechanisms, neither automatic

• Identity: reference-to-video - Pass the same reference images on every clip, address them by number, and generate a character sheet once. This is the only documented mechanism for keeping a subject consistent.
• Frames: image-to-video - Extract the final frame of clip N and pass it as the first frame of clip N+1. Pixel-level handoff depends on generation latency never jittering; the largest independent operator called it possible but fragile.
• Story: standing system prompt - fal.live narrative continuity comes from an operator-written system prompt plus viewer votes. Outside channels get the same effect with an LLM that remembers recent scenes and reference images for faces.

Sound is two fields, not a parameter

• Name sources and materials in the soundscape; a thin soundscape invites hallucinated audio.
• Music only in non_diegetic_music, as instruments, tempo, and dynamics.
• After writing N/A, no music or mood words anywhere else in the prompt. Stray words summon a score.
• For voiceover, write “says in an off-screen voiceover” and state that the on-screen lips do not move.

What stops a stream

• A blocked request: keep a shelf of pre-rendered fallback clips and treat rejection as “play fallback, log, move on”.
• Latency jitter: keep at least two clips of buffer and size it from measured, not published, throughput.
• Faces in wide shots and hands doing fine work: cut closer or cut away.
• Clones: two similar characters become one. Give them distinct features, colors, and movement.
• Drift: repeat the style block verbatim in every prompt.

The knobs that matter for a channel

From fal OpenAPI schemas for the three endpoints, read Sep 3, 2026.