Model description
AfroNaijaOldStyle LoRA for ACE-Step 1.5
This is an experimental LoRA adapter for ACE-Step 1.5 (Base) trained on ~31.5 hours of curated Afrobeats and classic Nigerian music styles. This is a "First Try" experimental release.
> Note: This adapter was trained on the `base` variant of ACE-Step. It is highly recommended to use it with the base model, as using it with the `turbo` variant may require significant scaling adjustments.
Model Details
- Adapter Trigger Tag: `afroNaijaOldaStyle`
- Dataset Size: 31 hours 29 minutes (420+ curated tracks) (a good number from the ORIN: The Nigerian music benchmark dataset
- Music Captioning: Done using nvidia/music-flamingo-think-2601-hf
- Training Rank: 128 (Alpha 256)
- Training Kit: Side-Step (Prodigy Optimizer)
Usage
To use this adapter, load it into the
ACE-Step Gradio UI
- Disable Quantization: Set `INT8 Quantization` to `None` in the Service Configuration (PEFT adapters are currently incompatible with torchao quantization in ACE-Step).
- Load Adapter: Point the LoRA Path to the folder containing these files.
- Trigger: Use the keyword `afroNaijaOldaStyle` in your prompt to activate the specific style.
Recommended Inference Settings (Mandatory for Quality)
Because this was trained on the Base model, you must change the default DiT parameters in the UI:
- Inference Steps: 50 (Do not use 8)
- Shift: 1.0 (Do not use 3.0)
- LoRA Scale: 0.7 - 1.0
Disclaimer
This model is released under the cc-by-nc-sa-4.0 license, inheriting the terms of the underlying datasets. It is intended for experimental and research purposes. Please respect artistic integrity and cultural heritage when generating content.
Trigger words
You should use afroNaijaOldaStyle, to trigger audio generation. (yes including the comma before description of the track, as a result of the way I hardcoded the tag into the dataset)
Download model
Download them in the Files & versions tab.
