Skip to content

Latest commit

 

History

History
83 lines (60 loc) · 4 KB

ADV_PARAMETERS_GUIDE.md

File metadata and controls

83 lines (60 loc) · 4 KB

HeartMuLa - Advanced Parameters Guide

Disclaimer:

The Default values listed below are explicitly cited from the official HeartMuLa research paper. The paper does not specify absolute minimum or maximum limits. The adjustments and "recipes" are derived from standard Large Language Model (LLM) behavior and community experimentation. While HeartMuLa is based on Llama-3.2, results may vary depending on the specific model checkpoint.

The "Mellow Bias": Why your Techno sounds like Pop

Many users report that the model ignores aggressive tags (e.g., Techno, Metal, High Energy) and defaults to a "mellow" or "pop" sound.

The Likely Cause: The HeartMuLa training dataset was filtered using Audiobox-Aesthetic scores to ensure high fidelity.

  • Fact: Aesthetic filters are trained to prefer "clean" audio.
  • The Side Effect: In generative audio, "clean" often correlates with Pop, Acoustic, or Soft textures, while "distorted" or "noisy" genres (like Dubstep or Rock) can be penalized.
  • The Result: When the model is unsure, it drifts toward the "safe" aesthetic (Mellow/Pop).

To break this bias, you must adjust the Inference Parameters to force the model away from its "safe" center.

1. Classifier-Free Guidance (--cfg_scale)

The "Strictness" Slider. This parameter controls how strongly the model forces the audio to match your text prompt (Tags/Lyrics) versus its internal training distribution.

  • Paper Default: 1.5
  • How it works (General AI Logic):
    • Low (1.0 - 1.5): The model is "loosely" guided by your tags. It prioritizes the internal "aesthetic" bias (smooth/safe audio).
    • High (2.0 - 4.0): The model is "forced" to match your tags, even if the result is less "aesthetically safe."

Community Observation: Users have reported that the default 1.5 is often too weak for specific genres. Increasing this value has helped users generate genres like R&B that were previously generating as Pop.

Recommendation: If your genre is being ignored, increase this value. Start with 2.5 and go up to 4.0 if needed.

2. Temperature (--temperature)

The "Randomness" Slider. This controls the probability distribution for the next audio token.

  • Paper Default: 1.0
  • How it works (General AI Logic):
    • Lower (< 1.0): The model becomes "conservative." It picks only the most likely sounds. This usually results in more repetitive, structured, and coherent rhythms.
    • Higher (> 1.0): The model becomes "creative" and takes risks. This adds variety but increases the chance of chaos or the melody falling apart.

Recommendation: For genres requiring strict rhythm (Techno, House), try lowering this slightly to 0.8 or 0.9 to lock in the groove.

3. Top-K (--topk)

The "Vocabulary" Limit. This limits the pool of possible "next sounds" to the top K most likely options.

  • Paper Default: 50
  • How it works:
    • A standard setting for Llama-based models. Lowering this (e.g., to 30) can reduce "hallucinations" or random artifacts, but may make the audio sound dull.

🧪 Experimental "Recipes"

These settings are suggestions based on how Autoregressive Transformers generally respond to these parameters. They are not official presets.

A. The "Aggressive / Specific" Fix

Use this if the model is ignoring your Genre tags (e.g., Metal, Techno, Rap).

  • Logic: High CFG forces the genre; Lower Temp keeps the rhythm tight.
  • Command: --cfg_scale 3.0 --temperature 0.8

B. The "Creative / Jazz" Flow

Use this for genres that benefit from improvisation or loose timing.

  • Logic: Moderate CFG allows some freedom; Higher Temp encourages unique melodies.
  • Command: --cfg_scale 2.0 --temperature 1.1

C. The "Safe / High Fidelity" (Paper Default)

Use this for Pop, Ballads, or when audio quality is the priority.

  • Logic: Low CFG prioritizes the "Aesthetic" filter; Default Temp ensures standard variety.
  • Command: --cfg_scale 1.5 --temperature 1.0