ACEStep 1.0 Extend Guider BETA ⚡🅡🅞🅣🅘

The straightforward answer to "this generated song is 30 seconds and I need 60": this guider generates new material before and/or after an existing clip's latent, growing it in either direction while trying to stay coherent with what's already there.

Note the display name calls it out as ACE-Step 1.0, and it's flagged BETA - this is the older-generation extend path in ryanontheinside's pack, from before ACE-Step 1.5's track-based system arrived.

How it works

Same Guider pattern as everything else in this family: it wraps model, conditioning, and CFG into a GUIDER you feed into a SamplerCustomAdvanced node. What makes it an "extend" guider specifically is that it takes your existing source_latents and generates new frames attached to either end of it, rather than generating a track from scratch.

The inputs and outputs that matter

• model, positive, negative - the usual conditioning trio.
• cfg - default 1.0 here, which matters: unlike the base-model default of 7 you'll see on other guiders in this pack, 1.0 suggests this is tuned around a distilled/guidance-baked-in setup. Pushing it up will roughly double render time for a pass that likely doesn't need it - the same guidance-distillation trap that shows up across modern generative models generally, not just this one.
• source_latents - the clip you're extending.
• extend_left_time / extend_right_time - seconds of new material to generate before and after, defaulting to 0 left / 10 right (i.e. by default it only extends forward).

Output is a GUIDER.

How to install it

Via ComfyUI Manager, search RyanOnTheInside. Manually:

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

then restart. Needs an ACE-Step 1.0-compatible checkpoint loaded upstream - this node doesn't fetch weights.

Common issues & troubleshooting

It's beta, and 1.0-specific. If you're working with ACE-Step 1.5 track-based generation, this isn't the node you want - reach for ACEStep15NativeLegoGuider and its family instead. Mixing 1.0 and 1.5 latents across nodes won't work; they're different formats with different frame rates.

New material drifts from the original. Extending significantly (tens of seconds) gives the model more room to wander away from the original's style than a short extension does. If coherence matters, extend in smaller increments and check the result before extending further, rather than asking for a large chunk in one shot.

Seam between old and new is audible. That's expected on its own - this guider only generates the extension. Pair it with ACEStepAudioPostProcessor afterward to crossfade the join smooth.