A composer that watches your film
Score a clip with music written for what happens on screen, not a generic mood.
Show video → LLM → music. Frames turns the clip into stills, a vision LLM writes a timed brief from them, and the music node composes from that brief. The blind branch beside it is the control.
Saved result
60 BPM, D minor to D major, wondrous nocturnal mood; instrumental, no vocals; celesta, harp, strings. 0-1s: dark sustained strings, unlit lantern on still lake. At 1s, when a spark flares overhead, celesta strikes a glissando; 1-3s harp figures follow the flame to the lantern's top. At 3s, hush—silence as flame touches. At 4s, when the lantern ignites, strings swell into a major chord with harp arpeggios. 5-7s: celesta melody grows as lanterns light across the water. 8-10s: full swell as lanterns rise, ending cleanly at 10s.
MiniMax H3 Max Turbo films 10 silent seconds of a lantern on a still mountain lake at night: a firefly drifts down, the lantern lights, and a wave of lanterns lights and rises across the water. Frames pulls one still per second, and Elorian Glimpse reads them and writes a composer's brief with tempo, key, instruments and cue points ("at 4 s, when the lantern ignites, strings swell"). ElevenLabs Music composes from that brief and Soundtrack lays it on the film. A second branch scores the same film from a text description only, so you can hear what watching adds. All four of Glimpse's cue points land within about half a second of the picture, and the fitted cue tracks the asked 60 BPM. Flaws: ElevenLabs ignored the 10 s duration on the fitted call and returned 19 s, so the fitted version stops at 10 s while still at full level. The hush asked for at 3 s is only a small dip. Glimpse sees 1 fps stills, so it called the firefly a spark. The clips were re-encoded for the gallery. Every sampled frame was checked and is SFW (lanterns on a lake, no people).
Another run: Dominoes and a silver bellSee what the same workflow made with different inputs ↓
Frames and both scores checked; the film misses the brief · 2026-10-30 · Reported saved run: $0.46826.
Partial. Same graph, a different shot: a curving chain of coloured dominoes on a sunny kitchen table that should end by knocking over a small silver bell. The film does not do that. The front row knocks the bell over at about 3–4 s rather than the last domino, the back row stays standing, and a second bell shows up at the right edge later on. The composer describes what is actually on screen (bell at 4 s, rolling until 6 s, still by 7 s) and asks for 112 bpm in C major. The fitted score comes in at that tempo, with a hit near 4 s and a swell near 6.5 s, but it is loud from the first second, where the brief asked for sparse. The blind score builds steadily to about 7 s and then fades, without matching any on-screen event. Both are instrumental.
Instrumental, 10 s, no vocals, 112 BPM, bright C major, playful. Instruments: pizzicato strings, marimba, handbell, brushed percussion. 0-2 s: sparse plucks; at 2 s, when dominoes begin toppling left, ticking pizzicato starts; at 3 s, bell tips, soft chime; at 4 s, bell knocked over, percussive hit; 4-6 s: marimba run rolls with the tumbling bell along the falling bottom row; at 6 s, swell as bell rolls right; at 7 s, hush as bell settles upright, one chime; 7-10 s: gentle resolve, clean end at 10 s.
Inputs for this version
Shot
Bright daytime on a sunny wooden kitchen table. A long curving line of colourful wooden dominoes stands still for a moment, then the first domino is tipped and they topple one after another in a fast chain reaction around the curve, and the very last domino knocks over a small silver bell, which rings and rolls. No people, no hands visible. Static camera, slight high angle, one continuous shot.
Ask
Score this 10-second clip.
Blind brief (no frames)
An instrumental film score cue, exactly 10 seconds long, no vocals, for this scene: on a sunny kitchen table a curving line of colourful dominoes topples in a fast chain reaction and the last one knocks over a small silver bell.
Sampled models: minimax/h3-max-turbo, elorian/glimpse, elevenlabs/music.
Make it yours
- Open the graph — save first if the canvas already has your work. Undo brings yours back.
- Change Shot for a different film, or swap the film node for a 📹 Video input to score your own clip. Change Ask to ask for a different style, or set the number of frames to match the clip length. Edit Blind brief (no frames) to match, or delete the blind branch once you have seen the difference.
- Check the editor estimate, then run with your NanoGPT key. H3 Max Turbo 10 s at 480p $0.25 + Glimpse $0.018 + two ElevenLabs Music cues at $0.10 each. Scoring your own clip costs about $0.12. The saved run was $0.468.
- Read the brief against the frames, then play both versions and listen at each cue point. Check the music's length, since music models can ignore the requested duration.
Download workflow · Original sampled graph · Explore MCP tools · awesome-noodles