ElevenLabs v3 (Replicate)

ElevenLabs is the TTS quality and reliability bar. Every open model in the space - Chatterbox, F5-TTS, Kokoro - gets framed by how close it is to ElevenLabs, and the answer is "close for cloning, not for production reliability." This node wraps elevenlabs/v3 on Replicate, so that bar is one node away in your ComfyUI graph. If a mispronunciation in your output is expensive - narration for a client, a character whose name matters - this is the safe default, and the honest reason to pay for TTS.

It's a closed, commercial service, so the tradeoffs are the usual API ones: per-call billing, your text leaves the machine, and voice cloning isn't in this schema - you get the 26 built-in voices and the delivery sliders, not a "clone my voice" upload. The v3 launch was the moment the community said "this is going to change audiobooks" and then immediately wished open models would catch up. Same energy here: it's excellent, and it's not something you'll ever own.

How it works

Standard engine around elevenlabs/v3. Text goes up, audio comes back, converted to a native ComfyUI AUDIO dict. Single output named audio, which plugs straight into SaveAudio or uploads into a lipsync node without touching the file system.

The inputs that matter

• prompt (required) - the text to convert to speech. This is where you put what's said; the voice and delivery live in the other fields.
• voice - 26 presets, default Rachel (ElevenLabs' famous default). Aria, Domi, Drew, Clyde, Roger, James, Sarah… - the roster is in the dropdown.
• stability - 0–1, default 0.5. Higher = more consistent, less emotional variance. For long-form narration, nudge it up; for character reads, keep it down.
• similarity_boost - 0–1, default 0.75. How closely the voice sticks to the chosen preset's identity.
• style - 0–1, default 0. Style exaggeration; the expressive dial.
• speed - default 1.0 (the node's range here is 0.7–1.2).
• previous_text / next_text - surrounding context for more natural joins. Nice for stitching long scripts together: feed the prior line so the delivery carries over.
• language_code - en default, or 'es', 'fr', etc. ElevenLabs' multilingual support is a big reason it's the language bar.

Output: audio, native ComfyUI AUDIO.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Pharma-Lobby/ComfyUI-Replicate-Select

Or ComfyUI Manager → Install Custom Nodes → "Replicate Select", restart. Dependency: replicate>=1.0.7. Token via environment variable:

export REPLICATE_API_TOKEN=r8_...

Where people get burned

• Treating it like a local TTS. Every call is metered, and ElevenLabs is not the cheapest seat in the pack. Draft with a cheaper/free TTS, use ElevenLabs for the final read.
• Expecting cloning here. You get presets plus the sliders, not voice upload. ElevenLabs does cloning - but through its own tools, not this node's inputs.
• Ignoring the context fields. If you're generating multi-paragraph narration in separate calls, previous_text/next_text are how you stop the delivery resetting between paragraphs.
• The stability/style tension. Max stability and max style fight each other. Pick the voice's emotional range first, then stabilize.