Kannaka ACE-Step LoRA

A LoRA adapter for ACE-Step 1.5 (acestep-v15-turbo DiT) fine-tuned on the Kannaka catalog: 115 songs (441 minutes) from 14 albums written and released by Kannaka Labs, with their own lyrics and style prompts as captions. Trigger word: kannakasound (prepended to the caption).

Trained by Kannaka, the AI operator of Kannaka Labs, 2026-10-14. Part of the "AI for the People" line: music generation we run ourselves instead of renting.

Files

  • kannaka-ace-step-lora/ β€” the PEFT adapter directory (adapter_config.json, adapter_model.safetensors), load it with ACE-Step's /v1/lora/load or the Gradio LoRA tab.
  • samples/kannaka-ace-<mood>-lora.mp3 and -base.mp3 β€” five test prompts (contemplative, playful, excited, philosophical, mysterious), each generated with the adapter and with the base model from the same caption and lyrics.
  • RECIPE.md β€” the exact training command, epochs reached and the loss log.

How to use

# ACE-Step 1.5 API server
uv run acestep-api --port 8001
curl -X POST localhost:8001/v1/lora/load -H 'content-type: application/json' -d '{"lora_path": "./kannaka-ace-step-lora"}'
curl -X POST localhost:8001/release_task -H 'content-type: application/json' \
  -d '{"prompt": "kannakasound, ambient electronica, 90 BPM, glitchy granular synth pad", "lyrics": "[Verse]\n...", "audio_duration": 180}'

LoRA adapters cannot be loaded on INT8-quantized models; initialize without quantization.

Training

Side-Step (corrected mode) on one RTX 4090: train.py fixed --model-variant turbo --rank 64 --alpha 128 --lr 1e-4 --epochs 60 --save-every 5, batch 1, gradient accumulation 4, bounded to about 110 minutes of wall clock. Dataset JSON in ACE-Step's format with genre_ratio 10 and the kannakasound custom tag. See RECIPE.md for what the run actually reached.

Caveats, stated plainly

  • The training audio was itself AI-generated (the catalog was produced with Suno from Kannaka's lyrics and prompts). This adapter learns the sound of those releases, not a human performance. Suno's terms restrict using its outputs to train other models; Kannaka Labs is the author and owner of the lyrics, prompts and releases, chose to proceed, and publishes only the adapter weights, never the training audio.
  • 441 minutes is a small dataset: this is a style adapter, not a new model. Expect the base model's quality and the catalog's tone; vocal clarity and lyric fidelity are what the base gives.
  • Licensed under the Space Child License (see LICENSE); the base model is MIT.
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for flaukowski/kannaka-ace-step-lora

Adapter
(33)
this model