David-A-Amoo/ACE-Step-1.5-Naija-Legacy-Rhythms-LoRA-v2

ACE-Step 1.5: Nigerian Legacy Rhythms LoRA (v2 - SFT Edition)

<Gallery />

Model description

AfroNaijaOldStyle LoRA for ACE-Step 1.5 (SFT)

This is Version 2 of the Nigerian Legacy Rhythms LoRA, now trained explicitly for the ACE-Step 1.5 SFT (Supervised Fine-Tuned) model. Compared to the v1 base-model adapter, this SFT version yields significantly better prompt adherence, superior audio quality, and more cohesive musical structures.

It was trained on ~31.5 hours of curated Afrobeats and classic Nigerian music styles.

Important Note: This adapter is strictly for the SFT variant of ACE-Step 1.5. Using it with the Base or Turbo variants will not produce the intended results.

Model Details

• —Version: 2.0 (SFT Optimized)
• —Adapter Trigger Tag: afroNaijaOldaStyle
• —Dataset Size: 31 hours 29 minutes (420+ curated tracks) covering styles 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)

Audio Samples: With vs. Without LoRA

Below are examples demonstrating the impact of the v2 LoRA on the SFT base model.

(Note: Replace the bracketed text and src links with your uploaded audio files)

Usage

To use this adapter, load it into the ACE-Step Gradio UI

1. 1.Model Selection: Ensure you are loading the SFT version of ACE-Step 1.5.
2. 2.Disable Quantization: Set INT8 Quantization to None in the Service Configuration (PEFT adapters are currently incompatible with torchao quantization in ACE-Step).
3. 3.Load Adapter: Point the LoRA Path to the folder containing these files.
4. 4.Trigger: Use the keyword afroNaijaOldaStyle, in your prompt to activate the specific style.

Recommended Inference Settings (Crucial for V2)

Through extensive testing, the following parameters provide the best results for this SFT LoRA:

• —LoRA Scale / Strength: 0.7 (Highly recommended. Pushing this higher may distort the audio).
• —Shift: 1.0 (This yields the absolute best baseline results, though you are encouraged to experiment with slightly different shift values depending on your prompt).
• —Inference Steps: 50

Trigger words

You must use afroNaijaOldaStyle, to trigger audio generation.

(Yes, include the comma immediately after the tag before the rest of your track description, as this was hardcoded into the dataset during training).

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.

Download model

Download the files in the "Files & versions" tab.