ACE-Step Simple Mode

A ComfyUI node in Audio/ACE-Step with 15 inputs and 4 outputs.

By kana112233·Created 6 months ago·Updated 6 months ago· 27

◄query►

◄checkpoint_dirAce-Step1.5►

◄config_pathacestep-v15-turbo►

◄lm_model_pathacestep-5Hz-lm-1.7B►

◄batch_size2►

◄seed-1►

◄inference_steps8►

◄deviceauto►

◄instrumentalfalse►

◄vocal_languageauto►

◄quantizationNone►

◄compile_modelfalse►

◄thinkingtrue►

◄audio_formatflac►

CategoryAudio/ACE-Step

Inputs (15)

Name	Type	Default	Description
query	STRING		Natural language description or prompt for music generation.
checkpoint_dir	COMBO	Ace-Step1.5	Directory containing ACE-Step model weights (DiT model).
config_path	COMBO	acestep-v15-turbo	Specific model configuration to use (e.g., v1.5 turbo).
lm_model_path	COMBO	acestep-5Hz-lm-1.7B	Path to the language model used for generating lyrics and metadata.
batch_size	INT	21–8	Number of audio samples to generate in a single batch.
seed	INT	-1-1–18446744073709550000	Random seed for reproducibility. Set to -1 for random generation.
inference_steps	INT	81–64	Number of diffusion steps. Higher values (e.g., 25-50) improve quality but are slower.
device	COMBO	auto	Computing platform to run the model on.
instrumentalopt	BOOLEAN	false	Whether to generate instrumental music only (no vocals).
vocal_languageopt	COMBO	auto	Vocal language (e.g., zh, en, ja).
quantizationopt	COMBO	None	Model quantization (e.g., int8). Reduces VRAM usage but requires torchao and compile_model=True. Incompatible with LoRA.
compile_modelopt	BOOLEAN	false	Whether to use torch.compile to optimize the model. Required for quantization. Slow on first run but faster afterwards.
thinkingopt	BOOLEAN	true	Whether to show the language model's Chain-of-Thought reasoning.
audio_formatopt	COMBO	flac	Output audio file format.
lora_infoopt	ACE_STEP_LORA_INFO		Optional LoRA model information for style fine-tuning.