ACE-Step 1.5 Silence Latent ⚡🅡🅞🅣🅘

The name undersells it a little. This doesn't give you a clip of digital silence - it loads ACE-Step 1.5's own learned representation of an empty stretch of audio, in latent space. That's a subtly different thing, and it's useful for two reasons: setting an exact output duration before you've generated anything, and controlling how strongly a reference track's timbre bleeds into a new one.

Part of ryanontheinside's ACE-Step 1.5 toolset, alongside the track-based Lego Guider and the extend/repaint nodes.

How it works

ACE-Step works in latent frames, not raw seconds, so if you want a sampler to produce exactly 30 seconds of output, you need a correctly-sized starting latent to hand it - that's what this node builds. It's also, per the node's own description, meant to be blended with a reference_latent elsewhere in your graph: mixing in some of this learned-silence latent dilutes how much of the reference's timbre carries over, letting you dial the strength down instead of it being all-or-nothing.

The inputs and outputs that matter

• seconds - the only input. Default 30, range 0.1 to 400. The tooltip is explicit that this converts to latent frames at 25 fps - that's the frame rate ACE-Step 1.5 latents use, worth remembering if you're doing any manual frame math elsewhere.

Output is a single LATENT, ready to plug in wherever a guider or blend step wants a starting or reference latent.

How to install it

Via ComfyUI Manager, search RyanOnTheInside. Manually:

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

then restart. No extra model download for this one - it's generating from the ACE-Step 1.5 model you've already loaded elsewhere in the graph.

Common issues & troubleshooting

Mixing up latent frame rates across versions. ACE-Step 1.5 nodes in this pack use 25 fps; the older 1.0-era nodes (like AudioLatentInfo) default to a different rate entirely (10.77 fps). If durations look wrong downstream, you're probably feeding a 1.5 latent into something expecting 1.0 timing, or vice versa - check which family of nodes you're actually working in.

Expecting real silence and getting something else. If you plug this straight into playback expecting a quiet gap, remember it's a latent, not decoded audio - it needs to go through a sampler or decode step like any other latent before it's actually sound, and depending on how it's used it may not decode to true silence at all, just the model's low-information baseline.