Nodes/ComfyUI-vlo/vlo Feather Audio Latent Mask
ComfyUI Node

vlo Feather Audio Latent Mask

Softens the time edges of an audio latent noise mask so inpainted audio joins the surrounding audio without a click. A fractional mask value is a genuine per-step denoise strength, not just a crossfade: MiniMax H3 places those rows at sigma = mask * sigma_audio and conditions them accordingly, so the model generates the transition. Run this after any blank latent composite, which leaves outer ramps with the original audio to blend toward. Inner and centered ramps sit inside the region that composite cleared, so those need original_audio_latent connected as well.

By PxTicks·Created 4 months ago·Updated a day ago· 0
vlo Feather Audio Latent Mask
  • audio_latent
  • original_audio_latent
  • audio_vae
  • audio_latent
◄modeouter►
◄lead_ramp0.15►
◄tail_ramp0.15►
◄lead_hold0.10►
◄tail_hold0.10►
◄curvecosine►
◄floor0.00►
◄layout_overrideauto►
◄audio_latent_rate0.00►
Categorylatent/audio

Inputs (12)

NameTypeDefaultDescription
audio_latentLATENTStandalone audio latent or nested AV latent that already carries an audio noise mask. Nested video masks are preserved.
modeCOMBOouterWhere the ramp sits relative to the masked region. 'outer' keeps the region solid and decays outward into the preserved audio, 'centered' straddles the edge, and 'inner' keeps the ramp inside so no preserved step is touched. The latter two shrink the region, so a region narrower than the ramp loses steps to it.
lead_rampFLOAT0.150–10Seconds of ramp before each masked region. Rounded to whole audio latent steps (25 ms for MiniMax, 40 ms for LTX); a non-zero value always gets at least one step.
tail_rampFLOAT0.150–10Seconds of ramp after each masked region. Worth lengthening past the lead when a note or phoneme decays for longer than it takes to start.
lead_holdFLOAT0.100–10Seconds of fully solid mask added before each region, ahead of the ramp. Audible onsets lead the visible motion that drew the mask, so a hold hands those steps to the model outright.
tail_holdFLOAT0.100–10Seconds of fully solid mask added after each region, ahead of the ramp. Useful when reverb or a decay tail outlasts the masked frames.
curveCOMBOcosineRamp shape. 'cosine' and 'smoothstep' flatten at both ends, so there is no kink where the ramp meets the solid core; 'linear' and 'exponential' leave one.
floorFLOAT0.000–1Minimum mask value everywhere, applied after the ramp. Above zero this lightly denoises the whole preserved region, which can even out a level or timbre step but rewrites audio you asked to keep. Zero leaves preserved audio untouched.
layout_overrideCOMBOautoAuto prefers latent/VAE metadata, then recognizes current LTX [B,C,T,F] and MiniMax [B,C,S,T] VAEs.
audio_latent_rateFLOAT0.000–1000Audio latent steps per second, used to convert the ramp and hold seconds. Zero resolves this from metadata or the VAE.
original_audio_latentoptLATENTThe same latent from before any blank latent composite. Its audio is restored underneath the ramp, so ramp steps blend toward the original audio instead of toward silence. Required for 'inner' and 'centered' after a composite, since those ramps sit inside the region it cleared; 'outer' ramps fall outside it and need nothing.
audio_vaeoptVAEAudio VAE used to resolve layout and latent rate automatically. It may be omitted when the latent carries metadata or overrides are supplied.

Outputs (1)

NameTypeDescription
audio_latentLATENT—