Doubao 音色克隆 / Seed ICL 2.0 语音合成

This node is two operations wearing one coat. On doubao-voice-clone it takes a short voice sample, trains a voice from it, and hands back a preview clip plus the ID you need later. On doubao-seed-icl-2.0 it takes that ID and makes the voice read whatever you type. Train, then speak. Getting that order right is basically the entire learning curve, and once it clicks this is one of the most immediately useful things in the pack.

Worth situating it: open TTS in ComfyUI is a real option - Kokoro for speed, Chatterbox for zero-shot cloning, F5-TTS for the fast path (audio-generation.md) - but those want a GPU and a download. This wants neither. It's ByteDance's Doubao voice stack behind an API call, and ByteDance keeps its flagship products hosted while the components ship free (bytedance.json), so there's no local version to wait for. The catch is the usual one: your sample and your script leave the machine.

What each mode actually does

In clone mode the payload carries a metadata.audio_url pointing at your uploaded WAV (or a public URL you paste in), plus optional custom_speaker_id, reference_text and language. The prompt in this mode is a 4–300 character preview line - the sentence the freshly trained voice says back to you. It is not a description of the voice. When the job finishes, the node reads the speaker ID out of the response and exposes it on the speaker output.

In ICL mode (doubao-seed-icl-2.0) speaker is required, and the metadata is a synthesis request: output format, sample_rate, speech_rate, loudness_rate, pitch, silence_duration, and a voice_model of seed-tts-2.0-standard or seed-tts-2.0-expressive. The prompt is now your actual script, 1–10000 characters.

The inputs you actually set

model decides which half of the node you're in, and the front-end hides the irrelevant fields when you switch - so clone-only fields like audio_url genuinely disappear in synthesis mode. That's not a bug, it's the reason the node doesn't look like a form from the tax office.

audio_url and the optional reference_audio input are exactly one or the other. Both filled, or neither, and validation stops you before submission. A connected AUDIO gets encoded to WAV and rejected above 10 MB; a pasted URL must be plain HTTP(S). The tooltip notes raw PCM needs 24 kHz mono.

language should match the recording - 0 Chinese, 1 English, or api_default to omit the code. The author validated this node with the English code on an English sample, and a mismatch is the quickest way to a strange-sounding clone.

custom_speaker_id is optional - 8–256 characters, must start with an English letter - and it's your friendly name for the voice. Don't confuse it with training_speaker_id, which hides behind advanced_mode and is the provider's ID. The tooltip is blunt about this because people mix them up: the handle you feed into a synthesis run comes from the speaker output, not from either field.

voice_model matters more than it looks: picking a specific model means context_texts is not sent. Want to instruct the read - "warm, slightly hushed, unhurried", up to 500 characters - and you have to leave voice_model on api_default. Keep advanced_mode off, too; it only reveals seventeen fields like bit rate, subtitle toggles, watermark and LaTeX parsing that you almost certainly don't want.

Outputs

audio is the decoded clip - 24 kHz out of a clone run - into Save Audio or a preview. audio_path is the local file, audio_url the temporary link, task_id the remote job, response the raw JSON. Then speaker, which is the whole point: wire it into a second copy of this node set to ICL.

Two nodes: clone → speaker → a second copy set to ICL, same API key both times, since the trained voice lives in that account. Pasting the ID by hand works, but wiring the output is the reliable path.

Install

Same pack as the rest - 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

That's requests>=2.28 and nothing else; no models to download. You still need a key (Seedance API Config node, SEEDANCE_API_KEY, or config/.env) and a restart.

Where people get burned

The clip is too big - the 10 MB encoded limit catches people who upload a two-minute sample when ten seconds of clean speech is plenty. The clone prompt is under four characters, or someone wrote a description there instead of the line to be spoken. A training run finishes but returns no speaker ("training succeeded without a speaker ID") - try a cleaner sample. And people paste custom_speaker_id into the synthesis node's speaker field and get nothing; that's not the handle.

Two practical notes. skip_error here returns one second of silence at 24 kHz so a batch keeps moving, so a silent output means a failed job, not a quiet voice. And cloning is a feature where the rules are yours to keep: get consent for any voice that isn't your own, because the model will happily make it say anything you type.