zhenzhen-qwen3-tts-lowprice

Voice is the one thing the local ecosystem still reliably hands to the cloud, and this node is the budget version of that move. Comfly_qwen3_tts_lowprice runs Alibaba's Qwen3 text-to-speech models through zhenzhen's domestic low-price shop (api.seedance.nz), turns your text into an actual AUDIO tensor inside ComfyUI, and bills you per character-ish instead of making you rent a GPU. If you're doing narration, character dialogue, or an audiobook read, it's a "good enough voice for pennies" option.

The "lowprice" in the name is the whole story: unlike the pack's FAL nodes, which take a per-node api_key, this one takes an api_config dictionary that comes from the pack's API Settings node (Comfly_seedance2_low_price_settings). That node carries the base URL and token for the domestic shop, which the README says runs about 75–80% of the overseas price. You can feed it one shared settings node and drive every low-price node in your workflow from the same key.

How it works

It's an async task: submit the request, poll the shop's task endpoint until it's done, download the audio, and hand it back as a real AUDIO output you can wire into a video's audio track or a save node. No model files, no VRAM, no local torch inference for the speech itself.

The inputs that matter

• model - qwen3-tts-flash (default) or qwen3-tts-instruct-flash. The instruct variant is the one that honors instructions - the plain flash model ignores them.
• prompt - the text to speak.
• voice - a voice name, Cherry by default. There's a catalog of voices on the shop side; you type the name as a string.
• language_type - Chinese, English, Japanese, Korean, German, French, Russian, Portuguese and more. Pick it or the prosody will be off.
• instructions - only used by the instruct model: "speak slowly, sound like a news anchor," that sort of thing.
• optimize_instructions - default on; lets the model clean up your instruction before speaking.

The seed input is a ComfyUI cache-control seed only - it's deliberately not sent to the API, which doesn't expose one. Fixed reuses the cached result; randomize forces a new run. Don't go hunting for reproducible voice takes via seed; that's not what this is.

Outputs: audio (the tensor - wire it to video or a save node), audio_url, task_id (handy for looking the job up in the shop's async-task page if something goes wrong), and response for debugging.

Installing it

Same story as every zhenzhen node: ComfyUI Manager, search "Comfyui-zhenzhen", or

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

then restart. No model downloads - the dependencies are plain HTTP/audio-stack stuff. You need a token from the domestic shop (https://api.seedance.nz) and a Comfly_seedance2_low_price_settings node wired into api_config.

Common issues

• Empty audio with no error - you usually have skip_error on, which swallows failures and returns a blank. Flip it off while debugging.
• Wrong language prosody - the language_type matters more than it looks like it should.
• Audio doesn't appear on your video - check that you're feeding the audio tensor into the right input; the audio_url alone won't sync anything.

It's not ElevenLabs-grade cloning and the README doesn't pretend it is - this is the "cheap, fast, good enough for a voiceover" lane, and for that it does the job. The audio-generation side of our KB makes the same call: cloud wins where reliability matters, and this node makes that call cheap.