Qwen3 TTS Clone with Prompt (Fast)

This is the generation half of the pack's two-step voice cloning. The Create Clone Prompt node turns a reference clip into a QWEN3TTS_FAST_PROMPT voice asset once; this node takes that asset plus a line of text and speaks it. Wire them together and you've decoupled "who's talking" from "what they say" - build the voice once, then generate as many lines as you like with it, each run skipping the reference-audio analysis entirely.

If that sounds like a workflow with a lot of moving parts, that's because it is, and it's worth it precisely when you have many lines for the same voice. The one-shot Voice Clone node is simpler and perfectly good for a single take. But for a character with twenty lines, or a narrator whose every paragraph is a separate run, this setup is the difference between re-processing the same clip twenty times and just... generating. Same voice every time, no drift, no repeat work.

How it works

Load a Base checkpoint (1.7B Base is the usual; 0.6B Base if you want it light) - the clone-prompt machinery lives on Base models, and so does this node's underlying call. The three inputs you'll actually set:

• clone_prompt - the QWEN3TTS_FAST_PROMPT object from Create Clone Prompt (Fast). This is a type-matched wire, so you can't misroute it; it only accepts the prompt output.
• text - what the cloned voice says. Multiline.
• language and seed - as usual (Auto handles English; ten explicit languages on offer).

The optional max_new_tokens, temperature, top_p, and repetition_penalty do the standard sampling job; remember that in this fast path temperature only steers the talker stage, so it's a subtle dial, not a personality switcher.

The output is a single 24kHz mono AUDIO clip. From there it feeds a preview/save node, or straight into an audio-driven avatar pipeline if you're making a talking head.

How to install it

Same pack, same install. ComfyUI Manager → search "Qwen3-TTS" → install → restart, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/YildirimMC/ComfyUI-Qwen3-TTS-Fast
pip install -r ComfyUI-Qwen3-TTS-Fast/requirements.txt

NVIDIA GPU required. Models auto-download into ComfyUI/models/TTS/ on first load.

Common issues

No clone_prompt to plug in - you need the Create Clone Prompt node upstream, fed with a reference clip and (ideally) its transcript. There's no built-in prompt picker; the prompt is generated from audio, always.

Voice comes out generic - the quality of the voice was fixed when the prompt was built. A bad transcript or a noisy reference at that stage can't be rescued here; rebuild the prompt.

Cut-off output with a console warning - max_new_tokens got clamped to max_seq_len - 256. The clone prompt content counts against the static cache too; raise max_seq_len on the Loader if your lines are long.

Cryptic error at generation - check you're on a Base checkpoint. CustomVoice and VoiceDesign checkpoints don't have the clone path, and the failure message won't always say that plainly.

For a multi-speaker script, the efficient shape is: one Create Clone Prompt per character, one of these per character, and a way to stitch the outputs together in order. That's the workflow this pair of nodes was built for.