Sonilo Video to Music

Here's the node that feels like cheating. SoniloVideoToMusic watches a video and composes music to match it - not "moody backing track for an unspecified mood," but music aligned to the footage's pacing, tempo, and emotional beats. Trailer scores, social-cut soundtracks, ambient beds for gameplay clips: you drop in the video, out comes a score that lines up.

It's a partner node, so all the analysis and generation happens on Sonilo's servers through Comfy's API proxy - your machine uploads the video and streams back the finished audio. No Sonilo account or key; you log in with your Comfy account (Settings → User) and spend prepaid credits (Settings → Credits). Pricing is per second of generated audio, matching its text-to-music sibling, and the node landed in core around April 2026.

The inputs

• video - the footage, up to 6 minutes long. That's the hard ceiling; the upload is handled for you.
• prompt - optional, and here's the counterintuitive part: the tooltip recommends leaving it empty for best quality. When empty, the model fully analyzes the video and composes from what it sees; a prompt only makes sense when you want to steer it ("upbeat and hopeful" or "dark orchestral tension"). If you've been filling it in out of habit, try blanking it.
• seed - present, and - same as SoniloTextToMusic - ignored by the service, kept only for graph consistency. Don't expect reproducible takes.

Output and wiring

One AUDIO output into Preview Audio or Save Audio (Advanced). While it works you'll get live progress text - the model reports the detected video duration and even streams back a suggested track title, which is a fun way to see what it thinks your footage is about. The generated audio arrives in chunks and is reassembled into a single clip.

The honest gotchas

• Six minutes is the wall. Longer footage gets rejected at upload. Cut your video before you send it, not after.
• You're paying per second of output. A full-length cut is a full-length bill. Score a short representative segment first, verify the direction, then commit.
• Empty prompt really is better for a faithful match. Prompting fights the video analysis; use text to steer mood, not to describe what the model can already see.
• It composes a piece, not a stem project - no separated parts to tweak. Adjust via prompt or re-run, then mix in a DAW if you need surgical edits.
• The analysis is only as good as the footage. A static webcam shot gives the model little to work with; dynamic cuts get much better-scored music.

Ships with ComfyUI core, nothing to install. The fastest path from footage to a synchronized score, and one of the few nodes where "do less" is literally the documented best practice.