AceMusic Load LoRA

AceMusic Load LoRA is the pack's "give the model a specialty" node. If you're familiar with LoRA in image-land - a small adapter file that steers a base checkpoint toward a specific style without touching the big weights - this is the same idea for ACE-Step. The base model is a generalist; a LoRA nudges it toward a particular flavor, like the pack's default example, a Chinese rap adapter. Drop one in and your generations inherit its character.

Inputs:

• model - the ACEMUSIC_MODEL from AceMusic Model Loader. LoRA loading always chains after the loader: Model Loader → Load LoRA → every generation node.
• lora_name_or_path - a string identifying the adapter. The default is ACE-Step/ACE-Step-v1-chinese-rap-LoRA, which is a Hugging Face repo id - the engine fetches it by name. You can also point it at a local path.
• lora_weight - 0 to 2, default 1. How hard the adapter's influence is applied. 0 effectively disables it, 1 is the standard strength, above 1 pushes harder.

One output: model again - same type, so you can daisy-chain another LoRA or feed any downstream node. The node calls pipeline.load_lora(name, weight) on your loaded model and returns it with the adapter baked in for everything that follows in the chain.

The honest caveat

The ACE-Step LoRA ecosystem is small and young compared to the image side. The pack's README mentions fine-tuned adapters like RapMachine and Lyric2Vocal, but in practice the well-known one is the Chinese rap LoRA that ships as the default - a very specific flavor. There's no Civitai-scale directory of ACE-Step LoRAs to browse yet, so this node is more "load the one you found" than "curate a collection." When you do find one, the wiring is trivial and the effect is dramatic: the same caption that gave you generic pop suddenly comes back with the adapter's voice.

Installing it

Via ComfyUI Manager, search "ComfyUI-AceMusic", or:

cd ComfyUI/custom_nodes
git clone https://github.com/hiroki-abe-58/ComfyUI-AceMusic.git
cd ComfyUI-AceMusic
pip install -r requirements.txt
pip install git+https://github.com/ace-step/ACE-Step.git

Then remember the adapter itself is downloaded from Hugging Face on first use - you need network access for that first load, just like the base model. The ACE-Step engine install is the one that breaks on pinned == versions; relax them in a clone or --no-deps the install.

Common issues

• "Couldn't find the LoRA" on first use - the adapter name or path is wrong, or the Hugging Face fetch failed. Double-check the repo id and your connection.
• No audible change after loading - lora_weight too low (try 1.0+), or the adapter isn't actually wired into the generation chain (check that every generator uses the output of the LoRA node, not the loader directly).
• Weight above 1 sounds broken - expected. 1.0–1.5 is the sane zone; above that adapters start to distort. Pull it back.
• Adapter works on the pack's nodes but not native ACE-Step nodes - the pack and Comfy-Org's native ACE-Step path load adapters differently; keep LoRA loading within one pack.

It's the smallest node in the pack and the one with the least to play with today. But it's the future-proofing: the moment a real ACE-Step LoRA scene develops, this is the node that plugs you in.