Qwen3-TTS 1.7B VoiceDesign

The VoiceDesign variant of Qwen3-TTS does something the Base node deliberately doesn't: it synthesizes a brand-new voice from a natural-language description instead of cloning an existing one. "A warm, cheerful male voice with moderate pace" is literally the example in the tooltip. No reference recording, no transcript, no similarity trade-offs - you describe the voice, and the model builds it. If you've ever needed a narrator who doesn't sound like the five voices every TTS product ships with, this is the node.

It's Alibaba's Qwen3-TTS, API-only (like the rest of the 3.x line), served through Runware. The "voice design" flavor is the newer, flashier mode - the community's launch reaction was genuine enthusiasm, with the usual caveat that it's a closed model you're renting by the call.

What you set

speech (the text to be spoken, via RUNWARE_SPEECH) and positivePrompt (the voice description) are both required. That's the whole job of the node - describe, speak, done.

The optional extras are few and honest:

• settings.maxNewTokens - audio output cap (default 2048). The tooltip's warning stands: pushing it way up risks the generation hanging.
• numberResults - generate variations, each a billed run.
• outputFormat - MP3/WAV/FLAC/OGG.
• ttl - URL lifetime for hosted results.

Output is audio (AUDIO dict, waveform + sample rate) → Save Audio. There's no transcript field and no xVectorOnly here, because there's no reference audio at all - this node is intentionally minimal, and that's a feature.

Installing

Part of the Runware/ComfyUI-Runware pack.

ComfyUI Manager: search Runware, install, restart.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

No model downloads (runware-sdk, pillow, soundfile). API key from runware.ai/api-keys (Settings panel or RUNWARE_API_KEY).

Where people get burned

Descriptions are the whole game, and vague ones give you a voice that sounds like every other TTS. Be specific about pitch range, pacing, character, and delivery ("a gravelly, unhurried narrator with a dry wit") - the model rewards concrete language. One practical note: since voice design has no reference, there's no guarantee of consistency between runs; if you need the same voice across a series of clips, generate one anchor line first and consider whether cloning that output via the Base node gives you more stability for the rest.