ACE-Step LoRA Loader

LoRAs in the image world let you bolt a style onto a checkpoint without retraining it. ACE-Step's LoRA Loader does the same thing for music - the pack's own example is a Chinese New Year LoRA that steers generations toward a festive sound. You pick a LoRA, set how strong you want it, and feed the result into any generation node. It's the pack's small node with big consequences, because music LoRAs are rarer and more valuable than image ones: you're not re-rolling a face, you're re-styling an entire sonic identity.

The catch is how picky the loader is about what counts as a LoRA. It uses PEFT, so a "LoRA" here is a directory containing an adapter_config.json - not a single .safetensors file. If you're used to dropping a lone file into models/loras, this is the first thing that'll trip you up.

How it works

The node scans two places and builds a dropdown: every folder directly under ComfyUI/models/loras/ and every folder under ComfyUI/models/Ace-Step1.5/loras/ - but only folders that contain adapter_config.json. Select one, set strength (0 to 5, default 1.0), and it emits a single lora_info wire of the pack's custom ACE_STEP_LORA_INFO type. That wire plugs into Text to Music, Cover, Simple Mode - anywhere the tooltip says "Optional LoRA style model." Internally it resolves the folder to an absolute path and hands the generation code a path + scale pair, which gets applied via PEFT when the model loads.

The strength dial behaves like image LoRA strength: 1.0 is the trained style at full weight, lower blends it toward the base model, higher pushes past what it was trained for (and usually starts to degrade).

The inputs that matter

Just two:

• lora_path - the dropdown of detected LoRA folders. It'll say "None" until you have at least one valid folder in place.
• strength - 1.0 default; 0.5 for subtle, 2.0+ for aggressive.

Output: lora_info → any ACE-Step generation node.

Install and setup

The node pack itself installs like the rest:

cd ComfyUI/custom_nodes
git clone https://github.com/kana112233/ComfyUI-kaola-ace-step.git
cd ComfyUI-kaola-ace-step && pip install -r requirements.txt

Then the actual task: get a PEFT-style LoRA. The README's steps are clear - put the LoRA folder (with adapter_config.json inside) in ComfyUI/models/loras/ or ComfyUI/models/Ace-Step1.5/loras/, then refresh the dropdown. peft is already in the pack's requirements, so the loader dependency is handled.

Where people get burned

Three things, in order of frequency:

1. The dropdown shows only "None". Either your folder lacks adapter_config.json, or it's in the wrong place - the scanner only looks in those two loras directories. A bare .safetensors file will never appear.
2. LoRA + quantization don't mix. The tooltip on quantization (in the generation nodes) says it outright: int8 quantization is incompatible with LoRA. Don't enable both.
3. ACE-Step LoRAs aren't interchangeable with image LoRAs. If someone hands you a file trained for SDXL, the loader won't even see it. Music LoRAs for ACE-Step are a young ecosystem - expect to wait for the good ones.