Doubao Seed Audio 1.0 音频生成

The party trick here is that you can hand this node a photo of a person and it will generate audio in that person's voice. Doubao Seed Audio 1.0 is ByteDance's closed audio model (Doubao is ByteDance's consumer AI brand; the Seed line never gets open weights), and this node is how it sneaks into your ComfyUI graph. It's not a text-to-speech toy - you can clone a voice from up to three reference clips, or let the model infer a voice from a single reference image, which is the workflow the pack's own example files are built around.

The name is a bit of a lie: nothing runs on your machine. This is an API-wrapper node from the T8mars/ComfyUI_Seedance pack, which fronts a budget AI gateway (default https://api.seedance.nz). You pay per call with an API key, not with VRAM.

How it works

The node hits the async /v1/audio/generations endpoint (not the sync /v1/audio/speech one - the README is explicit about that, because the model family does have both). Your prompt and any reference material get uploaded, a task ID comes back, the node polls until the job finishes, downloads the result, and decodes it into a ComfyUI AUDIO. Decoding doesn't even need torchaudio - it falls back to SciPy for the default 24 kHz stereo WAV, so a bare requests-only install just works.

The inputs that actually matter

• prompt - 5 to 2048 characters. This drives the content: narration text, singing direction, whatever the scene needs.
• speaker - an optional voice/ID from the API. Fine, but the interesting path is cloning.
• reference_image or reference_audio1..3 - here's the catch: these are mutually exclusive with speaker, and with each other. Pick one route: an image (the "identify this person" trick), or up to three reference clips (proper voice cloning). The tooltips spell it out, and the node will reject you if you try to mix sources.
• output_format - wav is the default and the README's advice: easiest for ComfyUI to decode. Switch to mp3/pcm/ogg_opus only if you need it downstream.
• speech_rate (−50 to 100), loudness_rate (−50 to 100), pitch_rate (−12 to 12) - your fine-control dials. Leave them at 0 to start.

Outputs are audio (the decoded AUDIO, wire it into any save/preview node), plus audio_url, audio_path, task_id and the raw response JSON for debugging.

Installing it

This ships inside T8mars/ComfyUI_Seedance - a small one-person API-wrapper pack. Install from ComfyUI Manager by searching "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 ComfyUI, then get a key from api.seedance.nz/console. Drop a Seedance API Config node in with base_url and api_key, and connect its api_config output here - or set SEEDANCE_API_KEY and skip the node.

Common issues

• "API key empty" - you skipped the config step; connect the Config node or set the env var.
• Weird WAV duration - the upstream sometimes stamps WAVs with a 0xFFFFFFFF streaming length marker, so strict readers report a wrong duration. The node itself decodes fine; if external software complains, re-save the audio through ComfyUI's audio save node.
• Node not in the menu after install - restart ComfyUI; check the plugin lives at ComfyUI/custom_nodes/ComfyUI_Seedance.
• Reference sources rejected - you wired speaker and a reference image. Pick one.

Remember the whole input leaves your machine for a server you don't control - that's the deal with any closed-model API node, and Seed Audio is ByteDance's model with ByteDance's moderation behind the gateway.