Qwen3-TTS Save Prompt

Qwen3-TTS Save Prompt is a small, boring node that does exactly one thing: it takes a voice embedding and writes it to disk so you don't have to recompute it ever again. If you've used the pack's Prompt Maker → Voice Clone path for a few sentences, you already know the pain this removes - every generation re-analyzes the reference clip. Save Prompt is the "make it permanent" button in that chain.

What it does

It takes a QWEN3_PROMPT (the output of Qwen3-TTS Prompt Maker) and a filename, and writes a .safetensors file into ComfyUI/models/Qwen3-TTS/prompts/. The filename defaults to voice_embedding; give it something descriptive like narrator_male or ava_clone. It strips an extension if you accidentally type one, so ava.safetensors and ava land at the same place.

The file it writes stores the speaker embedding plus metadata - whether it's an x-vector-only or in-context-learning prompt, the reference text if there was one, that sort of thing. All of it survives a restart, which is the point: your cloned voice is now an asset sitting in your models folder, not something living in a running graph.

The single filepath output (a string) is mainly there so you can wire the saved location into whatever tooling you use to track what's on disk. You can ignore it and it costs you nothing.

Why you'd use it

This node is the persistence half of the pack's voice-library workflow, and it's built to be paired with Qwen3-TTS Load Prompt on the other end:

• Save side: Voice Design or cloning → Prompt Maker → Save Prompt (once, when you've got an embedding you like)
• Reuse side: Load Prompt → Voice Clone (every session after, no reference audio needed)

On its own, Save Prompt is a utility with a single use. In that pair, it's what turns "I cloned a voice once" into "I have a reusable voice I can pull from a dropdown in any workflow." That's a genuinely nice property for a TTS pack, where re-running reference analysis on every sentence is the default tax.

Installing & troubleshooting

It comes with the rest of the pack: ComfyUI Manager (search "ComfyUI-Qwen3-TTS") or git clone https://github.com/DarioFT/ComfyUI-Qwen3-TTS into custom_nodes, then pip install -r requirements.txt manually - ComfyUI doesn't auto-install it. Watch the transformers==4.57.3 pin from qwen-tts.

Troubleshooting is almost embarrassingly simple. "Empty prompt - nothing to save" means you wired in a garbage/empty QWEN3_PROMPT, i.e. the Prompt Maker output wasn't valid. Want to overwrite an existing saved voice? Same filename, it just rewrites it. That's the whole node - no VRAM concern, no hang risk, it's just a write to disk. Sometimes the boring nodes are the ones that never break.