ACE-Step Simple Mode

If the rest of the ACE-Step pack feels like a production studio, Simple Mode is the jukebox with a text box. You type "energetic K-pop dance track with catchy hooks" and you get a finished song - no separate planning step, no fiddly caption/lyrics/BPM wiring. It's the node for "I just want to hear what the model can do," and it's also the best first node to click after you finish installing the pack, because it validates your whole setup in one run.

The trade-off is spelled out in the name. The multi-node path (Create Sample → Text to Music) lets you review and edit the plan before spending GPU time; Simple Mode collapses that into a single step, running the language-model planning pass and the diffusion generation back to back. You don't see or control the intermediate metadata - which is fine when you're iterating, annoying when a track is almost right and you want to tweak just the key.

How it works

Under the hood it's the same two-stage pipeline as the rest of the pack: the ACE-Step language model turns your query into lyrics and metadata, then the DiT renders the audio. Because that planning happens automatically, Simple Mode is where the thinking toggle shows its value - leave it on and the console shows the LLM's chain of thought, which is genuinely useful for figuring out why a track came out the way it did.

Outputs: audio (→ Save Audio), audio_path, metadata, and sample_info - that last one is the language model's full generated sample spec, the same "plan" you'd get from Create Sample, surfaced after the fact instead of before.

The inputs that matter

• query - the whole prompt. Same specificity advice as Text to Music: genre, mood, instrumentation, vibe.
• config_path / inference_steps - turbo at 8 steps by default; switch to acestep-v15-base with 32-64 for higher quality.
• batch_size - variations per run, up to 8.
• instrumental, vocal_language - the usual vocal controls.
• lora_info - accepts the pack's LoRA Loader output if you're applying a style LoRA.

The quantization/compile_model pair exists here too: int8_weight_only needs torchao and compile_model on, and won't combine with LoRA. Ignore unless you're VRAM-starved.

Install

Same shared setup as the whole Kaola pack:

git clone https://github.com/ACE-Step/ACE-Step-1.5.git
cd ACE-Step-1.5 && pip install -e .
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
huggingface-cli download ACE-Step/Ace-Step1.5 --local-dir ComfyUI/models/Ace-Step1.5

ComfyUI Manager installs the node pack; ACE-Step itself and the models are manual. Restart afterward.

Where people get burned

The usual pack-wide traps apply: "model path not found" means the weights aren't in ComfyUI/models/Ace-Step1.5/ with the right subfolders, and OOM means drop batch_size to 1 (and switch to the 0.6B language model). Because Simple Mode hides the plan, the "track's almost right but the key is off" situation is more annoying here than anywhere else - you can't edit metadata you never saw. That's the moment to rebuild the workflow with Create Sample + Text to Music instead of fighting the jukebox.