Qwen3-TTS 1.7B Base

Qwen3-TTS 1.7B Base is Alibaba's voice-cloning model, and this node is its ComfyUI front-end. You hand it a short reference clip of a voice and the text to speak, and it reads the text back in that voice. Alibaba's own marketing claims 3-second cloning across ten languages, and the community reaction at launch was genuinely warm - with the running complaint that, like most of Alibaba's newer models, it's API-only (the "can't count to 3 in open weights" joke gets a workout here). So this node isn't an alternative to a local TTS setup; it's the way you get at a model you can't download.

Two things set it apart from naive voice-clone tools. First, the input audio is a STRING - a Runware-hosted UUID or URL - not a ComfyUI AUDIO socket, because the audio never has to come through your graph to reach the API. Second, it gives you a real choice about how the clone works, via settings.xVectorOnly.

What you set

speech (the text to say, via a RUNWARE_SPEECH input) and audio (reference clip) are both required. The knobs:

• settings.xVectorOnly - the big one. true uses only a speaker embedding from the reference clip: simple, no transcript needed, lower similarity. false (default) uses ICL mode - the model learns from the reference audio's actual speech patterns, which needs a settings.transcript of that reference audio and gives noticeably better fidelity. If your reference clip is clean speech, write the transcript and use ICL.
• settings.maxNewTokens - audio output token cap (default 2048). Higher = longer output but the tooltip warns it "risks hangs." Don't crank it to infinity.
• settings.transcript - required for ICL mode, optional otherwise.
• numberResults - variations, each a billed run.

Output is audio as an AUDIO dict (waveform + sample rate) → Save Audio. outputFormat MP3/WAV/FLAC/OGG.

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

Deps are runware-sdk, pillow, soundfile; no model files. Runware API key from runware.ai/api-keys (Settings panel or RUNWARE_API_KEY).

Where people get burned

The xVectorOnly trap is the classic one: people skip the transcript, leave ICL mode on, and the clone sounds off - then blame the model. If you won't write the transcript, flip xVectorOnly on instead; it's the correct setting for a lazy-but-decent clone. Also, because the reference audio is a URL/UUID, a local file on your disk won't work - host it or upload via Runware first. And remember this is a voice clone: it's great for narration and characters, and you should be sensible about whose voice you're cloning and what you do with it.