ElevenlabsTextToEffect

The sibling node to ElevenlabsTextToSpeech, and honestly the more interesting one. You've rendered the video, animated the character, synced the lips - and now you need a whoosh, a crackle, rain on a window, a monster roar. This node turns a plain text description into 1–22 seconds of generated sound effect using ElevenLabs' text-to-SFX API. There's no local equivalent worth mentioning; generative sound effects simply aren't something the ComfyUI ecosystem does offline yet. If you want bespoke SFX inside the graph, this is the only realistic door.

How it works

Same architecture as the TTS node in this pack: it hits ElevenLabs' text_to_sound_effects.convert endpoint with your prompt, streams the MP3 bytes back, writes them to a temp file, decodes to float32 with soundfile, collapses to mono, and returns the standard ComfyUI audio dict {"waveform": tensor, "sample_rate": 44100}. Cloud-only, every run is a fresh API call, and every call costs credits. Don't queue a batch of twenty of these and walk away - that's a real bill.

The inputs that matter

The required pair is just text (multiline - describe the sound: "short electric buzz", "thunderclap with rain", "rustling fabric") and api_key. That's the whole skeleton; everything else is optional:

• duration (1–22 s, default 3): how long the effect runs. The 22-second cap is a hard API limit, so long ambient beds are off the table. 3–6 seconds is the sweet spot for most one-shot SFX.
• prompt_influence (0–1, default 0.3): how literally the text steers the result. This is the subtle one - a low value gives the model room to improvise and produce more organic, surprising sounds; cranking it to 1 tends to give you something rigid and literal. The 0.3 default is genuinely well chosen; most people who push it to max end up dialing it back.

That's the entire tuning surface. No voice settings, no model picker, and no exposed output format - the code just uses mp3_44100_128 internally, so don't go looking for format options in the node.

Wiring the output

The single audio output uses the same AUDIO dict as the TTS node, which means it drops into the same downstream slots: VHS audio nodes if you want to mix or save it, a lipsync or video-mux node if you want the effect to ride along with your clip. A practical pattern is generating the SFX here, a voice with the TTS node, and letting a video node merge both into the final render.

Installing it

Same pack as the TTS node - install once, get both. ComfyUI Manager: search "ElevenLabs". Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/sysL-padawan/comfyui-elevenlabs-integration.git
cd comfyui-elevenlabs-integration
pip install -r requirements.txt

Restart ComfyUI afterward. Dependencies are just elevenlabs and soundfile - no model downloads, no VRAM, no offline anything.

Gotchas and troubleshooting

This node has no local failure modes worth worrying about; the things that break are the things you'd expect from a cloud API. Invalid or unfunded key → auth error and a dead node. No network → dead node. Missing soundfile because you skipped the pip install → import error at load. The real trap is economic: short effects might feel cheap per call, but they add up fast when you're iterating on prompts, and the 3-second default means you're often paying for silence plus effect. Play with prompt_influence before you blame the quality - it's the one knob that changes the results more than anything else, and it's the easiest one to get wrong. For a zero-impression, barely-known pack, it's a remarkably clean little node: it does one API call and returns a usable tensor. Treat it as a paid convenience, not a pipeline you rely on offline, and it'll behave.