Teaching Generative Audio Models Non-Western Rhythms: Fine-Tuning ACE-Step 1.5 with LoRA on 6/8 South Indian Percussion

Description and use

Fine‑tuning the SouthIndiaParaiLoRA LoRA adapter on ACE‑Step 1.5 enables realistic 6/8 South Indian percussion generation, reducing loss by 75.5 %.

Problem

State‑of‑the‑art music generation models like ACE‑Step 1.5 are biased toward Western 4/4 time signatures, producing poor results for South Indian percussion such as Parai, Thappu, and Dappankuthu that rely on 6/8 compound rhythms and rapid transient spikes.

Outcome

The LoRA adapter reduces the final loss by 75.5% and enables realistic 6/8 South Indian percussion generation in ACE‑Step 1.5.

Workflow

• 1. Collect 393 studio‑quality stems of Parai, Thappu, Thappattam, and Urumi loops (80–172 BPM).
• 2. Train LoRA (rank 64, alpha 128) for 30 epochs on an NVIDIA A40 (48 GB VRAM) via RunPod.
• 3. Export adapter weights as .safetensors and adapter_config.json.
• 4. In ACE‑Step Playground, clone ACE‑Step‑1.5‑PlayGround‑Engine, drop the adapter files into checkpoints/SouthIndiaParaiLoRA/final/, launch WebUI, and select SouthIndiaParaiLoRA from the LoRA dropdown.
• 5. In ComfyUI, load adapter_model.safetensors into the ACE‑Step 1.5 LoRA loader node.
• 6. Prompt the model with a 6/8 South Indian rhythm description and set LoRA scale (e.g., 0.85).

Evidence

State‑of‑the‑art music generation models like ACE‑Step 1.5 are technical marvels, but they carry a known systemic blindspot: they were trained predominantly on Western 4/4 structures.

Source

To solve this, I fine‑tuned and open‑sourced SouthIndiaParaiLoRA on top of the ACE‑Step 1.5 foundation architecture.

Source

Slashed final loss by 75.5% (from a 1.97 baseline down to 0.48).

Source

The model weights are released in standard .safetensors format.

Source