You're looking at a specific version of this model. Jump to the model overview.

meta-innovation /heart_mula:f33e3f45

Input schema

The fields you can use to run this model with an API. If you don’t give a value for a field its default value will be used.

Field Type Default value Description
lyrics
string
[Verse] Every day the light returns Every day the fire burns [Chorus] Moving to the same steady beat
Lyrics for the song (text). Use sections like [Verse], [Chorus], [Bridge]. One line per phrase. Equivalent to --lyrics path content.
tags
string
piano,happy,acoustic,romantic
Style and mood tags, comma-separated without spaces. Examples: piano,happy,wedding,synthesizer,romantic. Equivalent to --tags path content.
max_audio_length_ms
integer
240000

Min: 10000

Max: 300000

Maximum audio length in milliseconds. 240000 = 4 minutes.
topk
integer
50

Min: 1

Max: 100

Top-k sampling for generation. Higher = more diversity; typical 30–80.
temperature
number
1

Min: 0.1

Max: 2

Sampling temperature. Higher = more random, lower = more deterministic. Typical 0.8–1.2.
cfg_scale
number
1.5

Min: 1

Max: 5

Classifier-free guidance scale. Higher values follow tags more closely. 1.5 is a good default.
version
None
3B
HeartMuLa version: 3B or 7B. In this deployment the model is loaded once at startup (3B); 7B not released yet.
mula_device
string
cuda
Device for HeartMuLa (e.g. cuda, cuda:0). In this deployment fixed at startup.
codec_device
string
cuda
Device for HeartCodec (e.g. cuda, cuda:1). In this deployment fixed at startup.
mula_dtype
None
bf16
Inference dtype for HeartMuLa. bf16 default; fp16/fp32 also supported. Fixed at startup.
codec_dtype
None
fp32
Inference dtype for HeartCodec. fp32 recommended; bf16 may reduce quality. Fixed at startup.
lazy_load
boolean
True
Lazy loading: load modules on demand to save GPU memory. Recommended for single GPU.

Output schema

The shape of the response you’ll get when you run this model with an API.

Schema
{'format': 'uri', 'title': 'Output', 'type': 'string'}