---
language:
- zu
tags:
- music-generation
- lora
- ace-step
- zulu
- maskandi
- amapiano
- isicathamiya
- south-african-music
base_model: ACE-Step/Ace-Step1.5
license: apache-2.0
---

# ACE-Step 1.5 — Zulu Music LoRA

A LoRA adapter fine-tuned on **63 Zulu music tracks** across traditional and modern South African genres, trained on top of [ACE-Step 1.5](https://github.com/ace-step/ACE-Step).

## Genres Covered
- Maskandi
- Amapiano
- Isicathamiya
- Kwaito
- Gqom
- Mbaqanga

## Training Details
| Parameter | Value |
|-----------|-------|
| Base model | ACE-Step 1.5 Turbo |
| Dataset | 63 Zulu songs |
| Epochs | 500 |
| Rank | 128 |
| Alpha | 256 |
| Batch size | 2 (effective 8) |
| Best loss | 0.4999 |
| Hardware | NVIDIA RTX 6000 Ada (48GB) |
| Language | Zulu (zu) |

## How to Use

Load the adapter in the ACE-Step 1.5 Gradio UI:

1. Launch the Gradio UI with `acestep-v15-turbo` as the main model
2. Under **🔧 LoRA Adapter**, enter the path to the `final/` folder
3. Click **Load LoRA** and enable **Use LoRA**
4. Set **LoRA Scale** to `1.0`
5. Generate with Zulu lyrics and a South African music caption

## Example Captions

**Maskandi:**
```
Maskandi guitar with Zulu male lead vocalist and choir in call-and-response harmonies, uplifting and nostalgic, 120 BPM, F major
```

**Amapiano:**
```
Amapiano track with Zulu male vocalist, log drum bassline, jazzy piano loops, handclaps, 114 BPM, A minor, energetic and joyful
```

**Isicathamiya:**
```
Traditional isicathamiya a cappella in Zulu with male lead vocalist and full choir, call-and-response structure, meditative and spiritual, 85 BPM
```

## Example Lyrics Format
```
[Verse 1]
Ngiyahamba emthunzini wobusuku
Inhliziyo yami iyazibuza
Ngabe indlela engiyihambayo
Iyabathokozisa yini abadala

[Chorus]
Bayangibona, bayangizwa
Noma ngithule, bayangizwa
Bayangibona, bayangivikela
Ngihamba negama labo

[Outro]
Hamba kahle
Hamba kahle
```

## Notes
- For best results use structured lyrics with section markers: `[Intro]`, `[Verse 1]`, `[Chorus]`, `[Bridge]`, `[Outro]`
- Disable INT8 Quantization before loading the LoRA adapter
- LoRA Scale between `0.8` and `1.0` gives the strongest Zulu style influence