arkennemasis fal · ElevenLabs Music · $0.6/min

There are three ElevenLabs music nodes in this pack and they cost exactly the same $0.60 per (rounded-up) minute, which raises the obvious question: why keep the oldest one? One reason. respect_sections_durations exists here and nowhere else, and it's the knob that decides whether a detailed composition plan is a specification or a suggestion.

If you're generating four bars of generic background music from a sentence, use Music v2 and move on. If you're scoring a cut to a structure - eight bars of setup, sixteen of build, a four-second sting on the reveal - this node lets you pin the section timings, and that's the difference between a plan you wrote and a plan the model agreed to.

How it works

The pack builds one node per fal endpoint from that endpoint's schema, so this node reflects the older fal-ai/elevenlabs/music endpoint rather than a newer model generation. On Run: key from .env, input checks, cost estimate against max_cost_usd, one billed submission, polling with live status, results downloaded into output/fal/fal-ai_elevenlabs_music/, request appended to the ledger.

The inputs that matter

• composition_plan - the JSON plan. Each section carries its own duration_ms, which is what respect_sections_durations acts on.
• respect_sections_durations - default true, which enforces every section's duration precisely. Set it false and the model may shuffle individual section lengths to land on the same total; the tooltip's own words are that this "generally lead[s] to better generation quality and improved latency". Translation: strict timing versus a better-sounding track. Pick your poison deliberately.
• prompt - the plain-language route. music_length_ms (3,000–600,000, -1 for auto) and force_instrumental both only apply on this path.

output_format defaults to mp3_44100_128 here - note that's lower than the v2 nodes' mp3_48000_192 default - and its option list is missing the 24 kHz and 48 kHz MP3 tiers that v2 offers. Worth knowing if you're matching a format your downstream tool expects. There's no seed widget on this node at all.

Outputs: audio, audio_path, info.

Money

$0.60 per output minute, rounded up. A useful consequence: a 4-second sting costs the same as a 59-second one, so batch your short cues into one generation if you can, or accept that a five-cue scene is $3.00 no matter how short the cues are.

Install

Manager → search comfyui-arkennemasis, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart ComfyUI and put FAL_KEY=your-key-from-fal.ai/dashboard/keys in the .env in the folder holding run_nvidia_gpu.bat (the node checks there and the ComfyUI folder, nowhere else). The fal nodes need no extra Python packages.

Common issues & troubleshooting

Your section timings came out wrong. Two candidates: respect_sections_durations set to false, or a plan whose duration_ms values don't add up to what you asked for in music_length_ms. Strict mode enforces what you wrote, so write it correctly.

"Composition plan" errors. It's parsed as JSON. A trailing comma or a smart quote pasted from a document will fail the input check before anything is sent - the good news is failures at that stage are free.

The output file sounds thin. You're probably on the default mp3_44100_128. Bump output_format; it's the same price.

Instant identical result. reuse_identical_run returned a previous paid run from output/fal. Flip it off for a fresh take.

Local alternative worth knowing. ACE-Step is the open model this community uses for music, good at instrumental beds, rougher on vocals, and free after electricity. If what you need is a bed, run it locally. Come here when you need the vocals to hold up.