Qwen3 TTS 语音合成（2 合 1）

Qwen3 TTS is Alibaba's text-to-speech line, and this node wraps its two modes: Flash (reads your text straight) and Instruct Flash (reads it and follows natural-language delivery instructions - "speak slowly and a little sad"). The Instruct mode is the reason this node is worth a look over a generic TTS call: you describe how to say it, not just what to say. Feed it the result_text from a prompt enhancer or narration script and you can have a voiced character line in a few calls.

As with the whole T8mars/ComfyUI_Seedance pack, it's an API wrapper - key from api.seedance.nz, requests only, nothing synthesized locally.

How it works

Standard async audio flow: submit the text with the chosen voice and language, poll, download, decode to ComfyUI AUDIO. The interesting part is the instructions field - for Instruct Flash, you can write a delivery direction in Chinese or English, and the optional optimize_instructions toggle refines non-empty instructions upstream before synthesis, which typically means less literal reading and more actual performance.

The inputs that matter

• model - Flash, or Instruct Flash.
• prompt - the text to synthesize.
• voice - a Qwen3 system voice ID or a custom voice ID you've set up.
• language_type - pick the language matching your text; don't make the model guess.
• instructions - Instruct Flash only: the delivery style. Ignored by plain Flash.
• optimize_instructions - Instruct Flash only: have the model polish your instruction first.

Outputs: audio (AUDIO), audio_url, audio_path, task_id, response.

Installing it

ComfyUI Manager → search "ComfyUI Seedance", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance.git
cd ../..
python -m pip install -r custom_nodes/ComfyUI_Seedance/requirements.txt

Restart, key from api.seedance.nz/console, Seedance API Config into api_config.

Common issues

• Instructions do nothing - you're on plain Flash; delivery directions are an Instruct Flash feature. Switch models.
• Wrong accent/pronunciation - language_type mismatching the text is the usual cause; set it to match the prompt.
• Long text feels slow - TTS jobs queue like everything else here; a long paragraph genuinely takes a while, and the node's poll timeout is generous for a reason.
• Voice quality varies by ID - system voices differ in warmth and stability; if one sounds robotic, try another before blaming your text. For a consistent character, the pack's MiniMax audio node gives you proper voice cloning.

Where this shines is batch narration: generate script lines, send each through, and keep the audio outputs lined up for a video's audio track. Keep prompts short and language-tagged and you'll get a surprising amount of mileage out of a very simple node.