🫨 Audio Latent Wobble (S42)

Now this one's fun, and it's also the rare audio node in this pack that works with machinery you already own. AudioLatentWobble takes an ordinary image or video latent - the kind any normal VAE produces - and pulses one of its channels in time with an audio curve. No audio VAE, no silent-latent traps, just beat-reactive distortion before your sampler runs. That's the whole trick: it doesn't touch audio at all, it uses an audio-derived curve to deform visual latents.

You feed it four things:

• latent - an image latent (4D B,C,H,W) or video latent (5D B,C,F,H,W) from any VAE encode.
• audio_curve - connect a bass, mid, or treble curve from S42_AudioFrequencyExtractor. Bass makes the kick drum drive the wobble; treble makes it twitchy and hi-hat driven. Your choice of band changes the feel completely.
• intensity (0–5, default 0.5) - how hard the beat multiplies the channel. The effect is channel *= 1 + (energy × intensity), so 0.5 is a gentle pulse and 3+ is where things visibly distort.
• channel_offset (0–15, default 0) - which latent channel gets hit. The tooltip's framing is right: different channels control different structural or color elements, so this is a "try each and see what moves" dial.

The mechanism adapts to your latent's shape, which is a genuinely nice touch. For a 4D image latent it takes the average energy across the whole clip and applies it once - your image latent pulses uniformly. For a 5D video latent it samples the curve per frame, so each frame gets the energy of the audio moment it lands on - that's the one that makes footage visibly breathe with the music.

Because this multiplies a channel in latent space before sampling, the result is not predictable like a filter. You won't be able to say "this channel controls brightness" with confidence - latent channels encode a messy mixture of semantics, and the tooltip admits as much. Expect to sweep channel_offset and find that one setting does something lovely and the rest do nothing visible. That's normal, not a bug.

It lives under the Experimental category for a reason. The pack author is clearly exploring audio-reactive generation, and this is the most usable fruit of that exploration: cheap, deterministic, and it doesn't require you to own any audio model beyond a waveform. It's the node I'd start a "beat-reactive music video" experiment with.

Installing it

Standard S42-CutFlow install: ComfyUI Manager → search "S42 CutFlow" → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/GeekyGhost/S42-CutFlow.git
pip install -r S42-CutFlow/requirements.txt

Only hard dependency is opencv-python-headless; the wobble math is pure torch. Look for the "[S42 CutFlow] Loaded ..." console line after restart.

Where you'll use it

Right after a VAE Encode, before the KSampler, with a Frequency Extractor's bass curve wired in. Pick a channel, pick an intensity, and let the kick drum push the latent around. It's experimental, results vary, and it's one of the few genuinely novel things in the pack.