Custom Voice (QwenTTS) Advanced

The basic Custom Voice (QwenTTS) node is great until it isn't. The moment your narration hums on past the final word, or you want it faster on a weak GPU, or you need a specific device/dtype - you want this one. Custom Voice (QwenTTS) Advanced is the same preset-speaker Qwen3-TTS generation, with every knob the basic node hides exposed.

What you get on top

The core job is identical: pick one of the nine preset speakers (Aiden, Dylan, Eric, Ono_Anna, Ryan, Serena, Sohee, Uncle_Fu, Vivian), type your text, get an AUDIO out. What the Advanced node adds is the sampler surface - and in this pack, that surface is exactly what fixes the two most-reported Qwen3-TTS problems.

The required inputs it adds over the basic node are device (auto/cpu) and precision (bf16/fp16/fp32). On a modern NVIDIA card, auto + bf16 is right. If you're CPU-only, drop to fp16 or fp32 - bf16 on CPU is slower than you'd think. Note the pack's trick for Apple Silicon: MPS forces fp16 regardless of what you pick, because that's what MPS likes.

The optional inputs that actually matter

• max_new_tokens (default 2048, range 256–4096) - this is the big one. Too high and short text generates trailing humming/noise; the README's own fix is 512–1024 for short lines. This is the "make it stop humming" control.
• do_sample (default false) - the pack's stability recommendation is to keep it off. Qwen3-TTS is much more consistent greedy.
• attention - auto/sage_attn/flash_attn/sdpa/eager. On CUDA, auto picks FlashAttention 2 if installed, then SageAttention, then plain SDPA. Both flash-attn and sageattention are optional CUDA-only installs - don't expect them to exist unless you pip-installed them:

  pip install flash-attn --no-build-isolation
  pip install sageattention   # experimental

• temperature, top_p, top_k, repetition_penalty - standard sampling knobs. Defaults (0.9/0.9/50/1.0) are sensible; bump repetition_penalty slightly if you hear the model getting stuck repeating syllables.
• seed - set it to reproduce a take; -1 for random.

Plus the same instruct style-instruction text and unload_models (default true) from the basic node.

Install and models

Same pack, same story as every node here: ComfyUI Manager or

cd ComfyUI/custom_nodes && git clone https://github.com/1038lab/ComfyUI-QwenTTS.git

then install requirements.txt with ComfyUI's python. The CustomVoice models (1.7B or 0.6B) auto-download to ComfyUI/models/TTS/Qwen3-TTS/ on first use - a few GB, so expect the first generation to be slow.

Gotchas

• The 'pad_token_id' transformers error means you're on a transformers 5.x nightly - pin transformers==4.57.3 and tokenizers<0.20.
• If you set attention=flash_attn but never installed flash-attn, the node silently falls back to SDPA. That's fine; just know the dropdown isn't a guarantee.
• Advanced nodes give you the power to make it worse. Keep do_sample=false and a sane max_new_tokens and you'll mostly avoid the "2-minute clip of stuttering" complaint people level at Qwen3-TTS clones - which, to be fair, is usually a clone-node problem, not a preset-voice problem.