ComfyuAudioNodes-BitsAndBobs
Nodes for ACE-Step 1.5 and XL: GGUF loaders, a multi LoRA/DoRA/LoKr loader with per-layer scaling, an experimental chord injector and a multi-audio comparison utility. Work in progress.
ComfyuAudioNodes-BitsAndBobs
[!WARNING] WORK IN PROGRESS (WIP)
This code is currently experimental, untuned, and not yet optimized. Use at your own risk. Expect breaking changes and potential bugs as development continues.
A collection of custom ComfyUI nodes for audio generation, comparison, and manipulation.
Nodes in this Collection
Lora-Dora-Lokr-Loader
A universal adapter loader for ACE-Step models.
- Supports LoRA, DoRA, and LoKr/LoHa (LyCORIS) formats.
- Features per-layer category scaling (Self-Attention, Cross-Attention, FFN).
- Advanced auto-strength balancing for Flux-based models.
- Includes a "Simple" node variant for a streamlined UI.
- Based on the DoRA Power LoRA Loader by xmarre.
Ace-Step_chord_injector
Tools for manipulating and injecting chord information into the ACE-Step generation pipeline.
[!NOTE] This node currently produces an audible effect on the output, but it is not yet performing its intended function correctly. It is included here for ongoing development and testing.
preview_audio_multi_compare
A utility node for side-by-side comparison of multiple audio generation outputs within the ComfyUI interface.
- Modified from components in the ryanontheinside repository.
ace_step_reference
A set of nodes for injecting reference audio into ACE-Step generation via multiple pathways.
- Timbre Encoding & Conditioning: Encodes reference audio into a timbre embedding and injects it into the cross-attention pathway. This method is stable and generally works well for transferring vocal/instrumental characteristics.
- KV Self-Attention Injection: Captures K/V tensors from a reference forward pass and injects them into the generation. This provides higher fidelity style transfer but is currently WIP (Work In Progress) with mixed results.
- Per-Step KV Injection: Real-time capture and injection at every sampling step. This is the most computationally expensive method but allows for precise alignment.
[!TIP] Timbre Conditioning and KV Injection can be chained inline to combine both methods for more comprehensive reference transfer.
See the included example workflow:
example_workflows/three_injection_methods.json.
- Includes a Reference Inspector debug node to verify capture output.
ace_step_gguf_loader
A custom GGUF and PyTorch bypass loader specifically designed for running quantized ACE-Step models natively inside ComfyUI.
- Supports ACE-Step 1.5 DiT
acesteparchitectures missing from standard allowlists. - Re-maps the GGUF
qwen3embedding namespace back into HuggingFace format for ComfyUI detection. - Includes a direct subclass wrapper for the
AudioOobleckVAEarchitecture to fix cross-device dtype crashes and apply missing 48kHz to 44.1kHz resampling when used with ACE-Step 1.5.
yue2_abc_manipulator
An all-in-one ABC score manipulator node for YuE2 audio generation workflows.
- Mode Selection:
bypass: Passes the ABC notation through untouched.instrumentalize: Moves Vocal melodies into resting Ins (instrumental) measures and mutes the Vocal notes.vocalize: Moves Ins melodies into resting Vocal measures using clean, minimal key-signature-aware accidentals and mutes the Ins notes.
- BPM Adjustment: Updates the score's declared tempo (
Q:field) to a target BPM or leaves it unchanged. - Time Signature Modification: Updates meter declarations (
M:field) to common signatures (2/4,3/4,4/4,5/4,6/4,6/8,7/8) or leaves them unchanged. - Pitch Transposition: Chromatically transposes notes up or down by up to two octaves (±24 semitones). Accurately handles octave wrapping across ABC register boundaries (uppercase/lowercase/octave marks) while preserving ties, rests, chord labels, and comments.
- Fully Self-Contained: Zero external dependencies outside the package.
yue2_lora_loader
A dedicated multi-adapter loader for YuE2 models, built on the Power-LoRA architecture.
- Supports LoRA, DoRA, LoKr, and LoHa adapter formats.
- Correctly targets both the AR/NAR diffusion transformer (
diffusion_model.model.layers.*) and the Qwen2 text encoder (text_encoders.model.layers.*). - Features a frontend picker UI with subfolder filtering to easily locate and stack multiple YuE2 adapters.
yue2_reference
A suite of reference audio injection and style transfer nodes designed for the YuE2-3B architecture (28 transformer layers, 48kHz AudioOobleckVAE).
- YuE2 Self-Attention Capture & Inject: Captures self-attention Key/Value activations from an encoded reference audio latent at a designated noise threshold and injects them into the generation pass across selected layer ranges.
- YuE2 Per-Step SA Inject (Per Layer): Performs real-time reference forward passes at each sampling step with granular per-layer strength sliders (layers 0–27), support for
inject,replace, andthresholdmodes, and layer taper options (linear,cosine). - YuE2 Per-Step SA Inject (step+time taper, step limited): Applies unified injection strength across layers with customizable step taper curves (
fade_out,cosine_fade_out,cosine_bell), time-based weighting, and step cutoffs to fade out reference guidance after a set step threshold.
Installation
Method 1: ComfyUI-Manager (Recommended)
- Search for
ComfyuAudioNodes-BitsAndBobsin the ComfyUI-Manager. - Click Install.
- Important: You must also install
ComfyUI-GGUFfrom the manager for the GGUF loader to function.
Method 2: Manual Installation
- Clone this repository into your ComfyUI
custom_nodesfolder:cd ComfyUI/custom_nodes git clone https://github.com/mmoalem/ComfyuAudioNodes-BitsAndBobs.git - Install the required Python dependencies. If you are using the ComfyUI Portable version, run this from your ComfyUI root folder:
For standard Python installs:.\python_embeded\python.exe -m pip install -r .\custom_nodes\ComfyuAudioNodes-BitsAndBobs\requirements.txtpip install -r requirements.txt - Mandatory Dependency: The GGUF loader requires the
ComfyUI-GGUFnodes to be present in yourcustom_nodesfolder:git clone https://github.com/city96/ComfyUI-GGUF - Restart ComfyUI.
Credits
Special thanks to the original authors whose work served as the foundation for these nodes:
- xmarre (DoRA Power LoRA Loader)
- ryanontheinside (RyanOnTheInside ComfyUI Nodes)
Support
If you find this project useful, consider supporting my work: