ACE-Step Model Sampling ⚡🅡🅞🅣🅘

This is the boring-looking node that quietly decides whether your ACE-Step generation sounds coherent or mushy. It's the ACE-Step equivalent of ModelSamplingAuraFlow or ModelSamplingSD3 - a one-parameter patch you drop between your model loader and your sampler to set the shift on the flow-matching schedule.

ACE-Step (the open-source text-to-music model that RyanOnTheInside has built an entire native pipeline around - see the ACE-Step 1.5 guiders and text-encode nodes in this same pack) is a flow-matching model, not a classic DDPM one. Flow-matching models predict a near-straight path from noise to signal, and how that path gets divided into steps is controlled by a shift value rather than a scheduler choice the way SD1.5/SDXL work. As the ComfyUI community has learned the hard way with other flow-matching models: shift is "roughly the flow-matching replacement for choosing a scheduler, and it is model-specific" - the right number for one model is wrong for another, and there's no substitute for checking what the model card or the pack author recommends.

How it works

You feed it your ACE-Step MODEL, it patches the sampling object with your chosen shift, and hands back a MODEL you plug straight into your sampler node. Nothing about your weights changes - this only changes how the sampler interprets timesteps during denoising.

The inputs and outputs that matter

• model (MODEL, required) - your loaded ACE-Step checkpoint, in.
• shift (FLOAT, default 1, range 0–100) - the only real knob. Higher shift pushes more of the sampling effort toward early, coarse structure; lower shift spends more steps on late detail. There's no universal "correct" value - it's tuned per model/checkpoint, so if ACE-Step's docs or a workflow you're copying specifies a shift, use that instead of the default.
• Output - MODEL, wired directly into your KSampler or the ACE-Step guider chain.

How to install it

Easiest path is ComfyUI Manager: search "RyanOnTheInside" and install. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/ryanontheinside/ComfyUI_RyanOnTheInside.git
cd ComfyUI_RyanOnTheInside
pip install -r requirements.txt

Restart ComfyUI afterward. This particular node doesn't pull any extra models itself, but if this is your first time touching the ACE-Step 1.5 nodes in this pack, be aware the guiders auto-download "silence latents" from HuggingFace on first use - so your first ACE-Step run needs internet access even if everything else is local.

Common issues & troubleshooting

"I changed shift and nothing sounds different." Two likely causes. First, make sure this node is actually upstream of the sampler in the graph - it's easy to wire your original model loader straight into the KSampler and leave this one dangling. Second, if you're chaining a custom scheduler downstream, some schedulers override the sigma shift you set here entirely and use their own - this is a documented gotcha on the image-generation side of ComfyUI's flow-matching nodes, and there's no reason to think ACE-Step's schedulers are exempt. If shift seems inert, check what scheduler you're actually using before concluding the node is broken.

Silent, garbled, or noisy output that isn't a shift problem at all. Before you touch this node, make sure the rest of your ACE-Step chain - text encode, guider, VAE - is wired correctly. Shift affects how the noise is removed, not whether your conditioning or task type is set up right, so it's rarely the first thing to debug when the output is flat wrong rather than just a bit off in character.

Update trouble via ComfyUI Manager. RyanOnTheInside's pack had a rough patch around its 2.0 release where Manager-based updates could get stuck. If you're on an old version and updates keep failing, a full uninstall and clean reinstall (rather than an in-place update) tends to fix it.