LoRA training with ACE-Step v1.5 (2B)
LoRA adapters are the fastest way to imprint a signature style on the ace step1.5 model. They let you fine-tune without retraining the full 2B parameter base model. The goal is to add a focused stylistic layer that you can enable or disable per project.
Data preparation
Keep the dataset narrow. A clean, consistent dataset yields a sharper LoRA than a large, mixed collection. Normalize your audio and organize files in a predictable structure so you can repeat the training later.
Tagging and prompts
Good tags make the LoRA usable. If you train for “vintage synthwave pad,” use that phrase consistently in your tags and prompts. The ACE Step 1.5 model will then respond to the phrase as a trigger for the style you trained.
Integration into your workflow
- Start with a low LoRA weight and increase gradually.
- Keep base prompts stable while testing your adapter.
- Save your best prompts and seeds for later reuse.