Qwen3 TTS Clone with Prompt

Qwen3 TTS Clone with Prompt is the consumer half of this pack's clone optimization. It doesn't take reference audio at all - it takes a ready-made QWEN_CLONE_PROMPT (built once by the Create Clone Prompt node) and just reads your text in that pre-baked voice. You use this node when you've already cloned a voice and you want the cheap part: many lines, same voice, no re-embedding.

The pitch is the same as Create Clone Prompt's, seen from the other end: the expensive work - turning your reference clip into a voice embedding - happened upstream, and this node skips it. For an audiobook or a multi-line narration in one voice, you chain Create Clone Prompt into a few of these and generation gets dramatically cheaper per line than re-running the plain clone node each time.

What it takes in

The required inputs:

• model - the QWEN_TTS_MODEL from the Model Loader (the Base model; it's cloning).
• text - what the voice says.
• clone_prompt - a QWEN_CLONE_PROMPT from the Create Clone Prompt node. Required, not optional - there's no fallback here, so wire it or the node won't run.
• language - Auto plus ten languages.
• seed - for reproducibility.

The optional sampling controls (max_new_tokens, temperature, top_p, repetition_penalty) are the usual defaults you can ignore until something sounds off. Output is a single AUDIO into Preview Audio / Save Audio.

Note the one input this node conspicuously doesn't have: no ref_audio, no ref_text, no x_vector_only. All of that got settled when the prompt was created. If you need to change the voice, change the reference upstream, not here.

How it works

It calls generate_voice_clone with voice_clone_prompt instead of raw reference audio - the prompt carries the speaker embedding the model already built, so the clone path is short. Same quality as a fresh clone (the prompt is the clone), just without the setup cost.

The honest shape of it

This node only makes sense as part of the two-node pair, and it's worth being clear: for a single one-off line, it's pointless - use the plain Voice Clone node. Where it shines is repeat work in the same voice. The standard long-form workflow looks like:

Model Loader (Base) → Create Clone Prompt → Clone with Prompt (line 1)
                                          → Clone with Prompt (line 2)
                                          → Clone with Prompt (line 3)

or feed the prompt into the Batch Generate node for the rest of the script. If you're doing a voiceover, a character script, or anything that's more than a couple of lines, this is the node you want generating them.

Install

One install for the whole pack: ComfyUI Manager search "Qwen3-TTS", or git clone https://github.com/PGCRT/ComfyUI-QWEN3_TTS into custom_nodes/, pip install -r requirements.txt, restart. The Base model auto-downloads into ComfyUI/models/TTS/ on first use. Keep the loader's keep_model_loaded on while you batch - every line will otherwise pay a model reload.