Audio
Sound in ComfyUI: loading and saving it, music from text, and video models that make their own sound.
In plain words
ComfyUI handles sound the way it handles pictures: an AUDIO value travels along links, audio models work on an audio latent, and an audio VAE turns it into a waveform. Sound appears in three kinds of graphs:
- Sound as an input: a voice recording drives a talking portrait.
- Sound from text: music or effects generated from a prompt.
- Sound with video: models that draw the clip and its sound together.
Where you meet it in a graph
- Load Audio reads a file from
input; Save Audio writes tooutput; Preview Audio plays it on the node. - Get Video Components gives a video's
audio; Create Video takesaudioand puts it into the clip. - Music and effects from text: ACE-Step (songs with lyrics) and Stable Audio Open (short sounds and loops) have built-in templates; they use an empty audio latent (for example EmptyLatentAudio or an ACE-Step latent node), a sampler, and VAE Decode Audio.
- Video with sound: LTX-2 (built-in templates) and Ovi (WanVideoWrapper) generate speech and sound with the frames.
- Speech to text, voice cloning, text to speech: through node packs; follow the pack's README, and the consent rule for voices.
Values and what they change
Length is set in seconds on the empty audio latent; longer sound costs more time. Music models read genre, instruments, mood and tempo from the prompt, and lyrics from a separate field.
Typical mistakes
- A clip saved without sound:
audionot connected to Create Video. - A frame rate on the video that does not match the speech: lips drift.
- Cloning a real person's voice without consent: see Content rules.
How the app helps
The lessons of the group Sound and speech build sound graphs: Music from text: ACE-Step 1.5, A sound from text: Stable Audio, Speech from text: Chatterbox, Sound for a silent video: MMAudio and Speech to text: Whisper. The talking-portrait lessons, the Ovi lesson and Video with sound: LTX-2.5 put sound into clips. On a node's card in the library, hold the pointer over the type mark of an AUDIO or VIDEO port to read the built-in note on that type.