Qwen3 TTS Create Clone Prompt

Qwen3 TTS Create Clone Prompt is the pack's optimization hack, and it's the one node whose name tells you exactly what it's for. Instead of re-running the full voice-clone pipeline every time you generate a line, you run it once on your reference audio, get back a reusable QWEN_CLONE_PROMPT, and hand that to the Clone with Prompt and Batch Generate nodes for every subsequent generation.

Why bother? Because a lot of the cost in cloning isn't the speech synthesis - it's building the voice embedding from the reference clip. If you're generating an audiobook, a podcast script, or ten tutorial lines in the same voice, re-cloning each line wastes that work every single time. The README claims 3–5x speedup on repeated generation, and even being skeptical of marketing numbers, the mechanism is real: one embedding built once, reused across runs.

Inputs and output

It's the simplest node in the pack:

• model - a QWEN_TTS_MODEL from the loader. Load the Base model (the cloning model); this is cloning, so CustomVoice/VoiceDesign models won't work.
• ref_audio - the reference clip as an AUDIO wire (from Load Audio or anywhere in the graph).
• ref_text - the exact transcript. Same rule as every clone node: get it right.
• x_vector_only (optional, default off) - build the prompt from just the speaker embedding and skip the transcript. Faster to build, lower fidelity.

The output is a single QWEN_CLONE_PROMPT. That's a custom type, so it only plugs into the two nodes that accept it - Clone with Prompt and Batch Generate - which keeps you from accidentally wiring it somewhere it doesn't belong.

The workflow it enables

Model Loader (Base) → Create Clone Prompt → Clone with Prompt (line 1)
                                          → Clone with Prompt (line 2)
                                          → Batch Generate (rest of script)

The QUICKSTART is explicit that this is the recommended path for anything long-form. The honest note: this is a convenience/perf node, not a quality node. It produces the same voice quality as the plain clone nodes - it just stops you from paying the embedding cost repeatedly. If you're only cloning one line ever, you don't need it. If you're cloning more than a couple of lines in the same voice, you absolutely do, and the graph stays tidier for it.

Install

Same pack, once: 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 downloads to ComfyUI/models/TTS/ on first use. Then: Load Audio → Create Clone Prompt → Clone with Prompt, and suddenly your "generate the whole script in this voice" workflow is no longer embarrassing to run.