AceMusic Model Loader

Every AceMusic workflow starts at the same place, and this is it. AceMusic Model Loader spins up the ACE-Step pipeline - a 3.5B parameter diffusion transformer trained by ACE Studio and StepFun - and hands you a single model handle you can wire into every other node in the pack. It's the closest thing to a checkpoint loader you'll find here, and like a checkpoint loader, it does the heavy lifting once and then mostly sits there.

The model isn't shipped with the pack. On first run the loader reaches out to Hugging Face and pulls ACE-Step-v1-3.5B down to ~/.cache/ace-step/checkpoints/ (or ComfyUI/models/acemusic/ if you've created that folder). That first load downloads a few gigabytes, so don't panic when the console looks dead for a while. Once it's cached, subsequent loads are fast.

The inputs that matter

• device - cuda by default, with cuda:0, cuda:1, and cpu available. Multi-GPU? Pick cuda:1 for the second card. CPU works but you'll be waiting - this is a 3.5B model.
• cpu_offload - the VRAM saver. Normal mode wants roughly 8GB of VRAM; offloading drops that toward ~4GB at the cost of speed. If you're on a 6GB card, this is your survival switch.
• torch_compile - a real speedup, but it requires Triton. On Windows that means pip install triton-windows first, and it only kicks in once you've installed it. Enable it only if you know Triton is present, or you'll trade a silent failure for a confusing one.
• overlapped_decode - faster generation for long clips (roughly anything over ~48 seconds). Cheap to leave on if you're making longer songs.

One output, model, typed ACEMUSIC_MODEL. It feeds every generation and editing node in the pack. Don't try to pass it to a normal checkpoint loader - this is a pack-private type.

How it works

Under the hood the node instantiates an ACEStepPipeline in bfloat16 on CUDA (float32 on CPU), then calls load_checkpoint(), which is what triggers the Hugging Face download. The loaded pipeline is cached by its settings - device, offload, compile, decode flags - so a workflow with multiple loaders pointing at the same config won't reload the model. That's a thoughtful touch, and it means you can leave a "spare" loader node around without paying for it.

ACE-Step is fast for a music diffusion model, which is part of why this pack is worth the setup. The README's numbers: a 4090 renders about a minute of audio in ~1.7 seconds, a 3090 in ~4.7s, and even an M2 Max does 26s for a minute of audio. You're not waiting minutes per take the way you do for images at high steps.

Installing it

Via ComfyUI Manager, search "ComfyUI-AceMusic". Or manually:

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

Then the important second step, installing the actual engine:

pip install git+https://github.com/ace-step/ACE-Step.git

The pack's own requirements are light - torch, torchaudio, soundfile, numpy, scipy - and notably it reads audio with soundfile/scipy instead of torchaudio's backends, which is the author's trick for keeping Windows installs from exploding. The friction lives in ACE-Step's own requirements.txt, which pins exact versions (torchaudio==2.10.0+cu128, matplotlib==3.10.1) that often don't exist for your Python. The README's fix: clone ACE-Step, relax == to >= in its requirements, and pip install -e . - or install the deps manually and pip install git+...ACE-Step.git --no-deps. This is the single most common failure point in the whole pack, so if the Model Loader throws an ImportError, that's where to look first.

Common issues

• ImportError: ACE-Step is not installed - you did the first install but not the second. Install ACE-Step.
• Download hangs or fails - check your connection to Hugging Face; you can also grab the checkpoint manually from https://huggingface.co/ACE-Step and drop it in the checkpoints folder.
• Out of VRAM - enable cpu_offload, or shorten your generation. The model's weights alone are several GB in bf16.
• torch_compile does nothing or errors on Windows - missing Triton; pip install triton-windows.

One output, one job. Load it once, feed it to everything, and the rest of the pack does the interesting stuff.