ComfyUI Node
Higgs v3 Generate
Generate Higgs Audio v3 speech without reference audio.
Higgs v3 Generate
- higgs_model
- audio
◄textHello! This is Higgs Audio v3 running natively inside ComfyUI.►
◄max_new_tokens2048►
◄temperature1.00►
◄top_p0.95►
◄top_k50►
◄seed0►
◄longform_chunkingtrue►
◄words_per_chunk45►
◄pause_between_chunks0.15►
CategoryHiggs v3 TTS
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| higgs_model | HIGGSV3TTS_MODEL | — | |
| text | STRING | Hello! This is Higgs Audio v3 running natively inside ComfyUI. | Text to synthesize. Inline tags work anywhere, for example <|emotion:relief|>, <|prosody:pause|>, or <|sfx:laughter|>Haha at the exact moment it should happen. |
| max_new_tokens | INT | 204832–8192 | Maximum audio-code tokens per single pass. 2048 is roughly 25-30 seconds; raise it or enable chunking if speech cuts off. |
| temperature | FLOAT | 1.000–2 | Sampling variety. 0 is greedy and repeatable; around 0.8-1.1 is usually natural. |
| top_p | FLOAT | 0.950–1 | Nucleus sampling cutoff. 1.0 disables it; 0.9-0.98 keeps speech expressive without wandering too much. |
| top_k | INT | 500–1026 | Limits each codebook sample to the top K choices. 0 disables; 50 is a steady default. |
| seed | INT | 00–2147483647 | 0 uses the current random state. A positive value is repeatable and is reused unchanged for every longform chunk. |
| longform_chunking | BOOLEAN | true | Split long text at sentence or pause-tag boundaries. Turn this on for narration; off is one direct pass and may stop early on long text. |
| words_per_chunk | INT | 4520–300 | Target words per chunk. Around 35-55 fits the 2048-token default better; raise with max_new_tokens for longer chunks. |
| pause_between_chunks | FLOAT | 0.150–2 | Seconds of silence inserted between longform chunks. Does not replace inline pause tags. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| audio | AUDIO | — |