Qwen3-TTS Voice Clone

Qwen3-TTS Voice Clone is the "make it sound like this specific person" node - the one you reach for when a preset voice won't cut it and you have a few seconds of someone talking. You feed it a reference clip, its transcription, and the text you want said, and it reproduces the voice. It's the hardest job in TTS and the one where open models have historically been weakest, so let me set expectations before you're disappointed.

The honest version: Qwen3-TTS is a solid clone, not a great one. The community consensus that's built up around this model line is that its likeness - how closely the output matches the target voice - trails Chatterbox and VibeVoice, and in blind comparisons people have called Qwen3-TTS clones "a random voice that happens to share the same pitch." For a narrator or a character-adjacent voice it's completely usable. For an impersonation that must hold up, it will disappoint. That's a model limitation, not a node bug, and it's worth knowing before you sink an afternoon into tuning it.

How it works

The node takes a QWEN3_MODEL from the Qwen3-TTS Loader - and you must load a Base variant (Qwen/Qwen3-TTS-12Hz-1.7B-Base or the 0.6B). The Base checkpoints are the only ones that carry the speaker encoder and cloning machinery. Then it needs two paths to a voice:

• ref_audio + ref_text - the direct way. Reference clip plus a transcription of what's in it. The transcription genuinely matters; feed it a transcript and likeness improves, skip it and the model has to guess.
• prompt (type QWEN3_PROMPT) - a precomputed voice embedding from the Qwen3-TTS Prompt Maker node. This is the fast path when you're generating many sentences with the same voice: compute the prompt once, reuse it, skip re-analyzing the reference clip every run.

Either one works; if you provide neither, the node raises a clear error. max_new_tokens defaults to 2048 and ref_audio_max_seconds to 30 - those two are your hang insurance (below).

The inputs that matter

• ref_audio - 1–10 seconds of clean, solo speech is the sweet spot. Longer isn't better; the model chokes on long or noisy references.
• ref_text - the exact transcription of that clip.
• text - what the cloned voice says. language is the same ten-language list as the pack's other generators, so a cloned voice can speak any of them.
• seed - 42 by default; change it when a run comes out weird.

The AUDIO output feeds any standard audio node, and from there into a lip-sync/avatar pipeline if that's the endgame.

Installing

ComfyUI Manager → search "ComfyUI-Qwen3-TTS", or git clone https://github.com/DarioFT/ComfyUI-Qwen3-TTS into custom_nodes and run pip install -r requirements.txt yourself (ComfyUI won't). Watch the transformers==4.57.3 pin from qwen-tts downgrading your environment - the pack README flags this, and it's the most common source of "my other nodes broke after I installed this."

Troubleshooting

This is where the community's pain is concentrated, and the pattern is consistent. Stuttering output - a clip that repeats the first word or spits out a long chain of stutters - is the signature failure. It's the known upstream Qwen3-TTS loop issue, and the fixes are exactly what the node's defaults are built around:

• Keep ref_audio short. The ref_audio_max_seconds parameter auto-trims long references (set to -1 to disable trimming, which you basically never want).
• Lower max_new_tokens from 2048 to ~1024 for short outputs. High token budgets invite the infinite loop.
• Change the seed. Some seeds are stable, some aren't.
• If the GPU is stuck at 100% and nothing returns, kill the Python process and restart ComfyUI.

If a clone is clean but just doesn't sound like the target, that's the likeness ceiling above - try a cleaner, shorter reference, and if you're chasing forensic-level accuracy you'll end up in Chatterbox or VibeVoice territory instead.