Learning Timeline
Key Insights
Step Count Optimization
For the 'Turbo model', a setting of 8 steps is optimal. Increasing the step count beyond that on a turbo model will result in diminishing returns, where the quality no longer improves significantly but takes much longer to process.
Quality & Creativity Tips
Use the 'SDE Inference Method' along with 'Chain of Thought' to improve song quality and coherence. If the results are too erratic, lower the 'LM temperature' to reduce randomness.
File Saving Warning
Generated songs are not automatically saved permanently. Ensure you manually click the 'Save' button; otherwise, the files may be lost once the application session is closed.
Prompts
Song Generation Prompt Example
Target:
Ace-Step 1.5
A pop song about Bitcoin crashing, female vocals, English
Step by Step
Ace-Step 1.5 Configuration and AI Song Generation
- Scroll down to the main music generation section after opening the app.
- Ignore the 'LoRA adapter' section for now unless you have a specific fine-tuned model for a particular style.
- Select your input mode: Use 'Default' to let the AI help build lyrics based on your description, or 'Custom' to manually enter your own 'Caption' and 'Lyrics'.
- Enter a description of the song in the input box (e.g., genre, mood, vocal type, and language).
- Toggle 'Only generate an instrumental song' to ON if you don't want any lyrics or vocals to be generated.
- Choose a language from the 'Language' dropdown. You can select 'Unknown' to allow the AI to auto-detect the language automatically.
- Set the technical parameters of the song, including 'BPM', 'Key', 'Time Signature', and 'Audio Duration'.
- Determine the number of songs you want to generate simultaneously in the 'Batch size' section.
- Click on 'Advanced Settings' to access in-depth technical configurations.
- Adjust the 'Step Count': Set it to 8 if you're using the 'Turbo model'. Use a higher value for the 'Base model'.
- Select your desired 'Audio Format'.
- Choose an 'Inference Method': Use 'Euler' for fast generation or 'SDE' for more coherent quality.
- Adjust 'LM temperature' for creativity levels (higher = more creative/random) and 'CFG scale' for prompt adherence (higher = follows the prompt more literally).
- Uncheck additional options at the bottom, such as 'Caption Rewrite', if you want to speed up the process by bypassing the language model.
- Click the 'Generate' button to start the generation process.
- Click the 'Save' button on the generated results, as files are only stored in a temporary folder by default.
