Audio Latent Blend (DEPRECATED USE FLEX LATENT BLEND) ⚡🅡🅞🅣🅘

Read the node's own display name before you spend time wiring it up: it's literally titled "Audio Latent Blend (DEPRECATED USE FLEX LATENT BLEND)". That's the author telling you, directly in the ComfyUI UI, not to build new workflows around this one. It still works - nothing here is broken - but it's a leftover from before the pack's "Flex" reactive system existed, and the newer Flex Latent Blend node in this same pack covers the same ground with the feature-modulation hooks this one never got.

What it actually does

If you land on this node anyway - maybe you opened someone else's older workflow and it's already wired in - here's the mechanism: it takes two LATENT tensors, latent1 and latent2, and blends them into one output latent using a blend_factor (0 to 1, how much of latent2 to mix in) and a blend_mode (normal, add, subtract, multiply, overlay). This is aimed at audio latents specifically - the kind you'd get out of an audio-generation model like ACE-Step - rather than image latents, though the math itself is generic tensor blending and doesn't actually care what the latent represents.

blend_mode changes what "blend" means: normal is a straight linear crossfade at whatever blend_factor you set, add and multiply combine the two latents' values directly (useful for layering rather than crossfading), subtract is closer to what you'd reach for to isolate a difference between two latents, and overlay applies the classic image-compositing overlay formula to latent space, which is a stranger, more contrast-heavy blend than the others.

Inputs and outputs

• latent1, latent2 (required, LATENT) - the two latents to blend.
• blend_factor (default 0.5, 0–1) - how much weight goes to latent2.
• blend_mode (default normal) - one of the five modes above.
• Output - a single blended LATENT, which continues into whatever decodes or samples further.

There's no feature/reactive input here at all - that's precisely the gap Flex Latent Blend fills, letting the blend factor itself react to audio or another driving signal over time instead of sitting at one fixed value for the whole clip.

Installing it

Install through ComfyUI Manager (search RyanOnTheInside), or:

cd ComfyUI/custom_nodes
git clone https://github.com/ryanontheinside/ComfyUI_RyanOnTheInside

Then pip install -r requirements.txt and restart ComfyUI. Both this node and its replacement ship in the same package, so there's no separate install step for the newer one - it's already there once the pack is installed.

Common issues

The only real "issue" here is using this node at all when you're starting fresh - go find Flex Latent Blend in the node search instead and you'll get the same blending plus the ability to drive blend_factor from an audio or motion feature. If you inherited a workflow that already has this node in it, it'll keep functioning; there's no indication the pack has actually removed the code, just steered new usage elsewhere.

Beyond that, this is a straightforward tensor op with no external model or heavy dependency, so if it fails, it's almost always upstream - a shape mismatch between latent1 and latent2 (different resolutions or channel counts) will error out rather than silently blend, which is the correct behavior even if the error message itself isn't always obvious about which side is wrong.