zhenzhen-doubao-seed-audio-1.0-lowprice

ByteDance's Seed Audio 1.0, as a ComfyUI node. Comfly_doubao_seed_audio_1_0_lowprice is a cloud TTS that takes text (or a reference voice/image) and returns an audio tensor you can wire straight into a video's audio track. It's the pack's answer to "my video needs a voice and I don't want to run Kokoro or Chatterbox locally."

The audio-generation KB's map is useful here: local TTS is a solved problem for generic narration, and voice cloning locally is where it gets fiddly - zero-shot clones need the right model and decent reference audio. Seed Audio is ByteDance's closed tier, resold through the Zhenzhen store at the domestic "lowprice" rate. The appeal is quality and instant voice cloning without the local TTS hobby.

The reference modes - the whole point

reference_mode is the personality switch, four options:

• none - plain text-to-speech. No reference needed.
• speaker - pick a pre-existing voice by the speaker string. The model has stock voices; you just name one.
• reference_audio - voice cloning. Feed 1–3 short audio clips (audio1–audio3, or reference_audio_urls) and the output speaks like them. This is the feature people actually reach for.
• reference_image - voice-from-face: a single image (reference_image / reference_image_url), and the model does its best to voice the person in it.

The validation is strict and helpful: none/speaker modes reject reference media; reference_audio needs 1–3 audios and forbids images; reference_image needs exactly one image and forbids audio. Get the mode wrong and it raises before anything is billed.

Inputs that matter

• prompt - the text to speak. Multiline.
• reference_mode - the four-way switch above.
• speaker - for speaker mode.
• output_format - wav / mp3 / ogg_opus.
• sample_rate - 8000 up to 44100, default 24000. 24k is the sensible TTS default; higher isn't automatically better.
• speech_rate (−50…+100), loudness_rate (−50…+100), pitch_rate (−12…+12) - the fine-tuning dials, 0 = neutral.
• api_config - zhenzhen-seedance2-low-price-api-settings, domestic key from api.seedance.nz.
• seed - cache-control only.

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

Install

ComfyUI Manager → search "Comfyui-zhenzhen" → install → restart, or:

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

No model downloads. Domestic-store key required.

Where people get burned

• The mode gates the inputs. Trying to clone a voice in speaker mode, or feeding an image in reference_audio mode, is a hard validation error. The mode is the contract - pick it first.
• Reference audio should be short and clean. The pack's TTS nodes elsewhere document upload rules (e.g. 10–240s for voice-clone materials in the Suno nodes); the same hygiene applies here. A noisy or overlong reference makes a mushy clone.
• It's a closed model through a reseller. Your text, reference audio, and face images all go to ByteDance's servers via the proxy. If you need privacy, local TTS (Kokoro, Chatterbox, F5-TTS) is the open alternative - this node is the convenience trade.
• Rate settings are per-millimeters, not presets. speech_rate isn't "fast/slow," it's a ±50 offset. 0 is the model's neutral; go in small steps.

This is the node for the "I need a professional voice, cloned, in my ComfyUI video, this afternoon" job. It won't replace the local TTS stack for people who care about privacy or fine control, but for turning a script into a track inside the same graph, it's hard to beat the convenience.